20210525.md

ArXiv cs.CV --Tue, 25 May 2021

1.SAT: 2D Semantics Assisted Training for 3D Visual Grounding ⬇️

3D visual grounding aims at grounding a natural language description about a 3D scene, usually represented in the form of 3D point clouds, to the targeted object region. Point clouds are sparse, noisy, and contain limited semantic information compared with 2D images. These inherent limitations make the 3D visual grounding problem more challenging. In this study, we propose 2D Semantics Assisted Training (SAT) that utilizes 2D image semantics in the training stage to ease point-cloud-language joint representation learning and assist 3D visual grounding. The main idea is to learn auxiliary alignments between rich, clean 2D object representations and the corresponding objects or mentioned entities in 3D scenes. SAT takes 2D object semantics, i.e., object label, image feature, and 2D geometric feature, as the extra input in training but does not require such inputs during inference. By effectively utilizing 2D semantics in training, our approach boosts the accuracy on the Nr3D dataset from 37.7% to 49.2%, which significantly surpasses the non-SAT baseline with the identical network architecture and inference input. Our approach outperforms the state of the art by large margins on multiple 3D visual grounding datasets, i.e., +10.4% absolute accuracy on Nr3D, +9.9% on Sr3D, and +5.6% on ScanRef.

2.luvHarris: A Practical Corner Detector for Event-cameras ⬇️

There have been a number of corner detection methods proposed for event cameras in the last years, since event-driven computer vision has become more accessible. Current state-of-the-art have either unsatisfactory accuracy or real-time performance when considered for practical use; random motion using a live camera in an unconstrained environment. In this paper, we present yet another method to perform corner detection, dubbed look-up event-Harris (luvHarris), that employs the Harris algorithm for high accuracy but manages an improved event throughput. Our method has two major contributions, 1. a novel "threshold ordinal event-surface" that removes certain tuning parameters and is well suited for Harris operations, and 2. an implementation of the Harris algorithm such that the computational load per-event is minimised and computational heavy convolutions are performed only 'as-fast-as-possible', i.e. only as computational resources are available. The result is a practical, real-time, and robust corner detector that runs more than $2.6\times$ the speed of current state-of-the-art; a necessity when using high-resolution event-camera in real-time. We explain the considerations taken for the approach, compare the algorithm to current state-of-the-art in terms of computational performance and detection accuracy, and discuss the validity of the proposed approach for event cameras.

3.Attention-guided Temporal Coherent Video Object Matting ⬇️

This paper proposes a novel deep learning-based video object matting method that can achieve temporally coherent matting results. Its key component is an attention-based temporal aggregation module that maximizes image matting networks' strength for video matting networks. This module computes temporal correlations for pixels adjacent to each other along the time axis in feature space to be robust against motion noises. We also design a novel loss term to train the attention weights, which drastically boosts the video matting performance. Besides, we show how to effectively solve the trimap generation problem by fine-tuning a state-of-the-art video object segmentation network with a sparse set of user-annotated keyframes. To facilitate video matting and trimap generation networks' training, we construct a large-scale video matting dataset with 80 training and 28 validation foreground video clips with ground-truth alpha mattes. Experimental results show that our method can generate high-quality alpha mattes for various videos featuring appearance change, occlusion, and fast motion. Our code and dataset can be found at this https URL

4.Multi-Level Attentive Convoluntional Neural Network for Crowd Counting ⬇️

Recently the crowd counting has received more and more attention. Especially the technology of high-density environment has become an important research content, and the relevant methods for the existence of extremely dense crowd are not optimal. In this paper, we propose a multi-level attentive Convolutional Neural Network (MLAttnCNN) for crowd counting. We extract high-level contextual information with multiple different scales applied in pooling, and use multi-level attention modules to enrich the characteristics at different layers to achieve more efficient multi-scale feature fusion, which is able to be used to generate a more accurate density map with dilated convolutions and a $1\times 1$ convolution. The extensive experiments on three available public datasets show that our proposed network achieves outperformance to the state-of-the-art approaches.

5.Large-Scale Attribute-Object Compositions ⬇️

We study the problem of learning how to predict attribute-object compositions from images, and its generalization to unseen compositions missing from the training data. To the best of our knowledge, this is a first large-scale study of this problem, involving hundreds of thousands of compositions. We train our framework with images from Instagram using hashtags as noisy weak supervision. We make careful design choices for data collection and modeling, in order to handle noisy annotations and unseen compositions. Finally, extensive evaluations show that learning to compose classifiers outperforms late fusion of individual attribute and object predictions, especially in the case of unseen attribute-object pairs.

6.Dynamic region proposal networks for semantic segmentation in automated glaucoma screening ⬇️

Screening for the diagnosis of glaucoma through a fundus image can be determined by the optic cup to disc diameter ratio (CDR), which requires the segmentation of the cup and disc regions. In this paper, we propose two novel approaches, namely Parameter-Shared Branched Network (PSBN) andWeak Region of Interest Model-based segmentation (WRoIM) to identify disc and cup boundaries. Unlike the previous approaches, the proposed methods are trained end-to-end through a single neural network architecture and use dynamic cropping instead of manual or traditional computer vision-based cropping. We are able to achieve similar performance as that of state-of-the-art approaches with less number of network parameters. Our experiments include comparison with different best known methods on publicly available Drishti-GS1 and RIM-ONE v3 datasets. With $7.8 \times 10^6$ parameters our approach achieves a Dice score of 0.96/0.89 for disc/cup segmentation on Drishti-GS1 data whereas the existing state-of-the-art approach uses $19.8\times 10^6$ parameters to achieve a dice score of 0.97/0.89.

7.Reconstructing Small 3D Objects in front of a Textured Background ⬇️

We present a technique for a complete 3D reconstruction of small objects moving in front of a textured background. It is a particular variation of multibody structure from motion, which specializes to two objects only. The scene is captured in several static configurations between which the relative pose of the two objects may change. We reconstruct every static configuration individually and segment the points locally by finding multiple poses of cameras that capture the scene's other configurations. Then, the local segmentation results are combined, and the reconstructions are merged into the resulting model of the scene. In experiments with real artifacts, we show that our approach has practical advantages when reconstructing 3D objects from all sides. In this setting, our method outperforms the state-of-the-art. We integrate our method into the state of the art 3D reconstruction pipeline COLMAP.

8.Multi-modal Understanding and Generation for Medical Images and Text via Vision-Language Pre-Training ⬇️

Recently a number of studies demonstrated impressive performance on diverse vision-language multi-modal tasks such as image captioning and visual question answering by extending the BERT architecture with multi-modal pre-training objectives. In this work we explore a broad set of multi-modal representation learning tasks in the medical domain, specifically using radiology images and the unstructured report. We propose Medical Vision Language Learner (MedViLL) which adopts a Transformer-based architecture combined with a novel multimodal attention masking scheme to maximize generalization performance for both vision-language understanding tasks (image-report retrieval, disease classification, medical visual question answering) and vision-language generation task (report generation). By rigorously evaluating the proposed model on four downstream tasks with two chest X-ray image datasets (MIMIC-CXR and Open-I), we empirically demonstrate the superior downstream task performance of MedViLL against various baselines including task-specific architectures.

9.Real-time Human Action Recognition Using Locally Aggregated Kinematic-Guided Skeletonlet and Supervised Hashing-by-Analysis Model ⬇️

3D action recognition is referred to as the classification of action sequences which consist of 3D skeleton joints. While many research work are devoted to 3D action recognition, it mainly suffers from three problems: highly complicated articulation, a great amount of noise, and a low implementation efficiency. To tackle all these problems, we propose a real-time 3D action recognition framework by integrating the locally aggregated kinematic-guided skeletonlet (LAKS) with a supervised hashing-by-analysis (SHA) model. We first define the skeletonlet as a few combinations of joint offsets grouped in terms of kinematic principle, and then represent an action sequence using LAKS, which consists of a denoising phase and a locally aggregating phase. The denoising phase detects the noisy action data and adjust it by replacing all the features within it with the features of the corresponding previous frame, while the locally aggregating phase sums the difference between an offset feature of the skeletonlet and its cluster center together over all the offset features of the sequence. Finally, the SHA model which combines sparse representation with a hashing model, aiming at promoting the recognition accuracy while maintaining a high efficiency. Experimental results on MSRAction3D, UTKinectAction3D and Florence3DAction datasets demonstrate that the proposed method outperforms state-of-the-art methods in both recognition accuracy and implementation efficiency.

10.LineCounter: Learning Handwritten Text Line Segmentation by Counting ⬇️

Handwritten Text Line Segmentation (HTLS) is a low-level but important task for many higher-level document processing tasks like handwritten text recognition. It is often formulated in terms of semantic segmentation or object detection in deep learning. However, both formulations have serious shortcomings. The former requires heavy post-processing of splitting/merging adjacent segments, while the latter may fail on dense or curved texts. In this paper, we propose a novel Line Counting formulation for HTLS -- that involves counting the number of text lines from the top at every pixel location. This formulation helps learn an end-to-end HTLS solution that directly predicts per-pixel line number for a given document image. Furthermore, we propose a deep neural network (DNN) model LineCounter to perform HTLS through the Line Counting formulation. Our extensive experiments on the three public datasets (ICDAR2013-HSC, HIT-MW, and VML-AHTE) demonstrate that LineCounter outperforms state-of-the-art HTLS approaches. Source code is available at this https URL.

11.WSSOD: A New Pipeline for Weakly- and Semi-Supervised Object Detection ⬇️

The performance of object detection, to a great extent, depends on the availability of large annotated datasets. To alleviate the annotation cost, the research community has explored a number of ways to exploit unlabeled or weakly labeled data. However, such efforts have met with limited success so far. In this work, we revisit the problem with a pragmatic standpoint, trying to explore a new balance between detection performance and annotation cost by jointly exploiting fully and weakly annotated data. Specifically, we propose a weakly- and semi-supervised object detection framework (WSSOD), which involves a two-stage learning procedure. An agent detector is first trained on a joint dataset and then used to predict pseudo bounding boxes on weakly-annotated images. The underlying assumptions in the current as well as common semi-supervised pipelines are also carefully examined under a unified EM formulation. On top of this framework, weakly-supervised loss (WSL), label attention and random pseudo-label sampling (RPS) strategies are introduced to relax these assumptions, bringing additional improvement on the efficacy of the detection pipeline. The proposed framework demonstrates remarkable performance on PASCAL-VOC and MSCOCO benchmark, achieving a high performance comparable to those obtained in fully-supervised settings, with only one third of the annotations.

12.What is the State of the Art of Computer Vision-Assisted Cytology? A Systematic Literature Review ⬇️

Cytology is a low-cost and non-invasive diagnostic procedure employed to support the diagnosis of a broad range of pathologies. Computer Vision technologies, by automatically generating quantitative and objective descriptions of examinations' contents, can help minimize the chances of misdiagnoses and shorten the time required for analysis. To identify the state-of-art of computer vision techniques currently applied to cytology, we conducted a Systematic Literature Review. We analyzed papers published in the last 5 years. The initial search was executed in September 2020 and resulted in 431 articles. After applying the inclusion/exclusion criteria, 157 papers remained, which we analyzed to build a picture of the tendencies and problems present in this research area, highlighting the computer vision methods, staining techniques, evaluation metrics, and the availability of the used datasets and computer code. As a result, we identified that the most used methods in the analyzed works are deep learning-based (70 papers), while fewer works employ classic computer vision only (101 papers). The most recurrent metric used for classification and object detection was the accuracy (33 papers and 5 papers), while for segmentation it was the Dice Similarity Coefficient (38 papers). Regarding staining techniques, Papanicolaou was the most employed one (130 papers), followed by H&E (20 papers) and Feulgen (5 papers). Twelve of the datasets used in the papers are publicly available, with the DTU/Herlev dataset being the most used one. We conclude that there still is a lack of high-quality datasets for many types of stains and most of the works are not mature enough to be applied in a daily clinical diagnostic routine. We also identified a growing tendency towards adopting deep learning-based approaches as the methods of choice.

13.SiamRCR: Reciprocal Classification and Regression for Visual Object Tracking ⬇️

Recently, most siamese network based trackers locate targets via object classification and bounding-box regression. Generally, they select the bounding-box with maximum classification confidence as the final prediction. This strategy may miss the right result due to the accuracy misalignment between classification and regression. In this paper, we propose a novel siamese tracking algorithm called SiamRCR, addressing this problem with a simple, light and effective solution. It builds reciprocal links between classification and regression branches, which can dynamically re-weight their losses for each positive sample. In addition, we add a localization branch to predict the localization accuracy, so that it can work as the replacement of the regression assistance link during inference. This branch makes the training and inference more consistent. Extensive experimental results demonstrate the effectiveness of SiamRCR and its superiority over the state-of-the-art competitors on GOT-10k, LaSOT, TrackingNet, OTB-2015, VOT-2018 and VOT-2019. Moreover, our SiamRCR runs at 65 FPS, far above the real-time requirement.

14.Towards Compact CNNs via Collaborative Compression ⬇️

Channel pruning and tensor decomposition have received extensive attention in convolutional neural network compression. However, these two techniques are traditionally deployed in an isolated manner, leading to significant accuracy drop when pursuing high compression rates. In this paper, we propose a Collaborative Compression (CC) scheme, which joints channel pruning and tensor decomposition to compress CNN models by simultaneously learning the model sparsity and low-rankness. Specifically, we first investigate the compression sensitivity of each layer in the network, and then propose a Global Compression Rate Optimization that transforms the decision problem of compression rate into an optimization problem. After that, we propose multi-step heuristic compression to remove redundant compression units step-by-step, which fully considers the effect of the remaining compression space (i.e., unremoved compression units). Our method demonstrates superior performance gains over previous ones on various datasets and backbone architectures. For example, we achieve 52.9% FLOPs reduction by removing 48.4% parameters on ResNet-50 with only a Top-1 accuracy drop of 0.56% on ImageNet 2012.

15.Mapping oil palm density at country scale: An active learning approach ⬇️

Accurate mapping of oil palm is important for understanding its past and future impact on the environment. We propose to map and count oil palms by estimating tree densities per pixel for large-scale analysis. This allows for fine-grained analysis, for example regarding different planting patterns. To that end, we propose a new, active deep learning method to estimate oil palm density at large scale from Sentinel-2 satellite images, and apply it to generate complete maps for Malaysia and Indonesia. What makes the regression of oil palm density challenging is the need for representative reference data that covers all relevant geographical conditions across a large territory. Specifically for density estimation, generating reference data involves counting individual trees. To keep the associated labelling effort low we propose an active learning (AL) approach that automatically chooses the most relevant samples to be labelled. Our method relies on estimates of the epistemic model uncertainty and of the diversity among samples, making it possible to retrieve an entire batch of relevant samples in a single iteration. Moreover, our algorithm has linear computational complexity and is easily parallelisable to cover large areas. We use our method to compute the first oil palm density map with $10,$m Ground Sampling Distance (GSD) , for all of Indonesia and Malaysia and for two different years, 2017 and 2019. The maps have a mean absolute error of $\pm$7.3 trees/$ha$, estimated from an independent validation set. We also analyse density variations between different states within a country and compare them to official estimates. According to our estimates there are, in total, $>1.2$ billion oil palms in Indonesia covering $>$15 million $ha$, and $>0.5$ billion oil palms in Malaysia covering $>6$ million $ha$.

16.Human-centric Relation Segmentation: Dataset and Solution ⬇️

Vision and language understanding techniques have achieved remarkable progress, but currently it is still difficult to well handle problems involving very fine-grained details. For example, when the robot is told to "bring me the book in the girl's left hand", most existing methods would fail if the girl holds one book respectively in her left and right hand. In this work, we introduce a new task named human-centric relation segmentation (HRS), as a fine-grained case of HOI-det. HRS aims to predict the relations between the human and surrounding entities and identify the relation-correlated human parts, which are represented as pixel-level masks. For the above exemplar case, our HRS task produces results in the form of relation triplets <girl [left hand], hold, book> and exacts segmentation masks of the book, with which the robot can easily accomplish the grabbing task. Correspondingly, we collect a new Person In Context (PIC) dataset for this new task, which contains 17,122 high-resolution images and densely annotated entity segmentation and relations, including 141 object categories, 23 relation categories and 25 semantic human parts. We also propose a Simultaneous Matching and Segmentation (SMS) framework as a solution to the HRS task. I Outputs of the three branches are fused to produce the final HRS results. Extensive experiments on PIC and V-COCO datasets show that the proposed SMS method outperforms baselines with the 36 FPS inference speed.

17.Out-of-Distribution Detection in Dermatology using Input Perturbation and Subset Scanning ⬇️

Recent advances in deep learning have led to breakthroughs in the development of automated skin disease classification. As we observe an increasing interest in these models in the dermatology space, it is crucial to address aspects such as the robustness towards input data distribution shifts. Current skin disease models could make incorrect inferences for test samples from different hardware devices and clinical settings or unknown disease samples, which are out-of-distribution (OOD) from the training this http URL this end, we propose a simple yet effective approach that detect these OOD samples prior to making any decision. The detection is performed via scanning in the latent space representation (e.g., activations of the inner layers of any pre-trained skin disease classifier). The input samples could also perturbed to maximise divergence of OOD samples. We validate our ODD detection approach in two use cases: 1) identify samples collected from different protocols, and 2) detect samples from unknown disease classes. Additionally, we evaluate the performance of the proposed approach and compare it with other state-of-the-art methods. Furthermore, data-driven dermatology applications may deepen the disparity in clinical care across racial and ethnic groups since most datasets are reported to suffer from bias in skin tone distribution. Therefore, we also evaluate the fairness of these OOD detection methods across different skin tones. Our experiments resulted in competitive performance across multiple datasets in detecting OOD samples, which could be used (in the future) to design more effective transfer learning techniques prior to inferring on these samples.

18.Face Anonymization by Manipulating Decoupled Identity Representation ⬇️

Privacy protection on human biological information has drawn increasing attention in recent years, among which face anonymization plays an importance role. We propose a novel approach which protects identity information of facial images from leakage with slightest modification. Specifically, we disentangle identity representation from other facial attributes leveraging the power of generative adversarial networks trained on a conditional multi-scale reconstruction (CMR) loss and an identity loss. We evaulate the disentangle ability of our model, and propose an effective method for identity anonymization, namely Anonymous Identity Generation (AIG), to reach the goal of face anonymization meanwhile maintaining similarity to the original image as much as possible. Quantitative and qualitative results demonstrate our method's superiority compared with the SOTAs on both visual quality and anonymization success rate.

19.Unsupervised Video Summarization with a Convolutional Attentive Adversarial Network ⬇️

With the explosive growth of video data, video summarization, which attempts to seek the minimum subset of frames while still conveying the main story, has become one of the hottest topics. Nowadays, substantial achievements have been made by supervised learning techniques, especially after the emergence of deep learning. However, it is extremely expensive and difficult to collect human annotation for large-scale video datasets. To address this problem, we propose a convolutional attentive adversarial network (CAAN), whose key idea is to build a deep summarizer in an unsupervised way. Upon the generative adversarial network, our overall framework consists of a generator and a discriminator. The former predicts importance scores for all frames of a video while the latter tries to distinguish the score-weighted frame features from original frame features. Specifically, the generator employs a fully convolutional sequence network to extract global representation of a video, and an attention-based network to output normalized importance scores. To learn the parameters, our objective function is composed of three loss functions, which can guide the frame-level importance score prediction collaboratively. To validate this proposed method, we have conducted extensive experiments on two public benchmarks SumMe and TVSum. The results show the superiority of our proposed method against other state-of-the-art unsupervised approaches. Our method even outperforms some published supervised approaches.

20.A Fourier-based Framework for Domain Generalization ⬇️

Modern deep neural networks suffer from performance degradation when evaluated on testing data under different distributions from training data. Domain generalization aims at tackling this problem by learning transferable knowledge from multiple source domains in order to generalize to unseen target domains. This paper introduces a novel Fourier-based perspective for domain generalization. The main assumption is that the Fourier phase information contains high-level semantics and is not easily affected by domain shifts. To force the model to capture phase information, we develop a novel Fourier-based data augmentation strategy called amplitude mix which linearly interpolates between the amplitude spectrums of two images. A dual-formed consistency loss called co-teacher regularization is further introduced between the predictions induced from original and augmented images. Extensive experiments on three benchmarks have demonstrated that the proposed method is able to achieve state-of-the-arts performance for domain generalization.

21.Dynamic Class Queue for Large Scale Face Recognition In the Wild ⬇️

Learning discriminative representation using large-scale face datasets in the wild is crucial for real-world applications, yet it remains challenging. The difficulties lie in many aspects and this work focus on computing resource constraint and long-tailed class distribution. Recently, classification-based representation learning with deep neural networks and well-designed losses have demonstrated good recognition performance. However, the computing and memory cost linearly scales up to the number of identities (classes) in the training set, and the learning process suffers from unbalanced classes. In this work, we propose a dynamic class queue (DCQ) to tackle these two problems. Specifically, for each iteration during training, a subset of classes for recognition are dynamically selected and their class weights are dynamically generated on-the-fly which are stored in a queue. Since only a subset of classes is selected for each iteration, the computing requirement is reduced. By using a single server without model parallel, we empirically verify in large-scale datasets that 10% of classes are sufficient to achieve similar performance as using all classes. Moreover, the class weights are dynamically generated in a few-shot manner and therefore suitable for tail classes with only a few instances. We show clear improvement over a strong baseline in the largest public dataset Megaface Challenge2 (MF2) which has 672K identities and over 88% of them have less than 10 instances. Code is available at this https URL

22.Oriented RepPoints for Aerial Object Detection ⬇️

In contrast to the oriented bounding boxes, point set representation has great potential to capture the detailed structure of instances with the arbitrary orientations, large aspect ratios and dense distribution in aerial images. However, the conventional point set-based approaches are handcrafted with the fixed locations using points-to-points supervision, which hurts their flexibility on the fine-grained feature extraction. To address these limitations, in this paper, we propose a novel approach to aerial object detection, named Oriented RepPoints. Specifically, we suggest to employ a set of adaptive points to capture the geometric and spatial information of the arbitrary-oriented objects, which is able to automatically arrange themselves over the object in a spatial and semantic scenario. To facilitate the supervised learning, the oriented conversion function is proposed to explicitly map the adaptive point set into an oriented bounding box. Moreover, we introduce an effective quality assessment measure to select the point set samples for training, which can choose the representative items with respect to their potentials on orientated object detection. Furthermore, we suggest a spatial constraint to penalize the outlier points outside the ground-truth bounding box. In addition to the traditional evaluation metric mAP focusing on overlap ratio, we propose a new metric mAOE to measure the orientation accuracy that is usually neglected in the previous studies on oriented object detection. Experiments on three widely used datasets including DOTA, HRSC2016 and UCAS-AOD demonstrate that our proposed approach is effective.

23.FineAction: A Fined Video Dataset for Temporal Action Localization ⬇️

On the existing benchmark datasets, THUMOS14 and ActivityNet, temporal action localization techniques have achieved great success. However, there are still existing some problems, such as the source of the action is too single, there are only sports categories in THUMOS14, coarse instances with uncertain boundaries in ActivityNet and HACS Segments interfering with proposal generation and behavior prediction. To take temporal action localization to a new level, we develop FineAction, a new large-scale fined video dataset collected from existing video datasets and web videos. Overall, this dataset contains 139K fined action instances densely annotated in almost 17K untrimmed videos spanning 106 action categories. FineAction has a more fined definition of action categories and high-quality annotations to reduce the boundary uncertainty compared to the existing action localization datasets. We systematically investigate representative methods of temporal action localization on our dataset and obtain some interesting findings with further analysis. Experimental results reveal that our FineAction brings new challenges for action localization on fined and multi-label instances with shorter duration. This dataset will be public in the future and we hope our FineAction could advance research towards temporal action localization. Our dataset website is at this https URL.

24.Towards Book Cover Design via Layout Graphs ⬇️

Book covers are intentionally designed and provide an introduction to a book. However, they typically require professional skills to design and produce the cover images. Thus, we propose a generative neural network that can produce book covers based on an easy-to-use layout graph. The layout graph contains objects such as text, natural scene objects, and solid color spaces. This layout graph is embedded using a graph convolutional neural network and then used with a mask proposal generator and a bounding-box generator and filled using an object proposal generator. Next, the objects are compiled into a single image and the entire network is trained using a combination of adversarial training, perceptual training, and reconstruction. Finally, a Style Retention Network (SRNet) is used to transfer the learned font style onto the desired text. Using the proposed method allows for easily controlled and unique book covers.

25.Recent Advances and Trends in Multimodal Deep Learning: A Review ⬇️

Deep Learning has implemented a wide range of applications and has become increasingly popular in recent years. The goal of multimodal deep learning is to create models that can process and link information using various modalities. Despite the extensive development made for unimodal learning, it still cannot cover all the aspects of human learning. Multimodal learning helps to understand and analyze better when various senses are engaged in the processing of information. This paper focuses on multiple types of modalities, i.e., image, video, text, audio, body gestures, facial expressions, and physiological signals. Detailed analysis of past and current baseline approaches and an in-depth study of recent advancements in multimodal deep learning applications has been provided. A fine-grained taxonomy of various multimodal deep learning applications is proposed, elaborating on different applications in more depth. Architectures and datasets used in these applications are also discussed, along with their evaluation metrics. Last, main issues are highlighted separately for each domain along with their possible future research directions.

26.Taylor saves for later: disentanglement for video prediction using Taylor representation ⬇️

Video prediction is a challenging task with wide application prospects in meteorology and robot systems. Existing works fail to trade off short-term and long-term prediction performances and extract robust latent dynamics laws in video frames. We propose a two-branch seq-to-seq deep model to disentangle the Taylor feature and the residual feature in video frames by a novel recurrent prediction module (TaylorCell) and residual module. TaylorCell can expand the video frames' high-dimensional features into the finite Taylor series to describe the latent laws. In TaylorCell, we propose the Taylor prediction unit (TPU) and the memory correction unit (MCU). TPU employs the first input frame's derivative information to predict the future frames, avoiding error accumulation. MCU distills all past frames' information to correct the predicted Taylor feature from TPU. Correspondingly, the residual module extracts the residual feature complementary to the Taylor feature. On three generalist datasets (Moving MNIST, TaxiBJ, Human 3.6), our model outperforms or reaches state-of-the-art models, and ablation experiments demonstrate the effectiveness of our model in long-term prediction.

27.High-level camera-LiDAR fusion for 3D object detection with machine learning ⬇️

This paper tackles the 3D object detection problem, which is of vital importance for applications such as autonomous driving. Our framework uses a Machine Learning (ML) pipeline on a combination of monocular camera and LiDAR data to detect vehicles in the surrounding 3D space of a moving platform. It uses frustum region proposals generated by State-Of-The-Art (SOTA) 2D object detectors to segment LiDAR point clouds into point clusters which represent potentially individual objects. We evaluate the performance of classical ML algorithms as part of an holistic pipeline for estimating the parameters of 3D bounding boxes which surround the vehicles around the moving platform. Our results demonstrate an efficient and accurate inference on a validation set, achieving an overall accuracy of 87.1%.

28.Deep Visual Anomaly detection with Negative Learning ⬇️

With the increase in the learning capability of deep convolution-based architectures, various applications of such models have been proposed over time. In the field of anomaly detection, improvements in deep learning opened new prospects of exploration for the researchers whom tried to automate the labor-intensive features of data collection. First, in terms of data collection, it is impossible to anticipate all the anomalies that might exist in a given environment. Second, assuming we limit the possibilities of anomalies, it will still be hard to record all these scenarios for the sake of training a model. Third, even if we manage to record a significant amount of abnormal data, it's laborious to annotate this data on pixel or even frame level. Various approaches address the problem by proposing one-class classification using generative models trained on only normal data. In such methods, only the normal data is used, which is abundantly available and doesn't require significant human input. However, these are trained with only normal data and at the test time, given abnormal data as input, may often generate normal-looking output. This happens due to the hallucination characteristic of generative models. Next, these systems are designed to not use abnormal examples during the training. In this paper, we propose anomaly detection with negative learning (ADNL), which employs the negative learning concept for the enhancement of anomaly detection by utilizing a very small number of labeled anomaly data as compared with the normal data during training. The idea is to limit the reconstruction capability of a generative model using the given a small amount of anomaly examples. This way, the network not only learns to reconstruct normal data but also encloses the normal distribution far from the possible distribution of anomalies.

29.Multi-Type-TD-TSR -- Extracting Tables from Document Images using a Multi-stage Pipeline for Table Detection and Table Structure Recognition: from OCR to Structured Table Representations ⬇️

As global trends are shifting towards data-driven industries, the demand for automated algorithms that can convert digital images of scanned documents into machine readable information is rapidly growing. Besides the opportunity of data digitization for the application of data analytic tools, there is also a massive improvement towards automation of processes, which previously would require manual inspection of the documents. Although the introduction of optical character recognition technologies mostly solved the task of converting human-readable characters from images into machine-readable characters, the task of extracting table semantics has been less focused on over the years. The recognition of tables consists of two main tasks, namely table detection and table structure recognition. Most prior work on this problem focuses on either task without offering an end-to-end solution or paying attention to real application conditions like rotated images or noise artefacts inside the document image. Recent work shows a clear trend towards deep learning approaches coupled with the use of transfer learning for the task of table structure recognition due to the lack of sufficiently large datasets. In this paper we present a multistage pipeline named Multi-Type-TD-TSR, which offers an end-to-end solution for the problem of table recognition. It utilizes state-of-the-art deep learning models for table detection and differentiates between 3 different types of tables based on the tables' borders. For the table structure recognition we use a deterministic non-data driven algorithm, which works on all table types. We additionally present two algorithms. One for unbordered tables and one for bordered tables, which are the base of the used table structure recognition algorithm. We evaluate Multi-Type-TD-TSR on the ICDAR 2019 table structure recognition dataset and achieve a new state-of-the-art.

30.Revisiting 2D Convolutional Neural Networks for Graph-based Applications ⬇️

Graph convolutional networks (GCNs) are widely used in graph-based applications such as graph classification and segmentation. However, current GCNs have limitations on implementation such as network architectures due to their irregular inputs. In contrast, convolutional neural networks (CNNs) are capable of extracting rich features from large-scale input data, but they do not support general graph inputs. To bridge the gap between GCNs and CNNs, in this paper we study the problem of how to effectively and efficiently map general graphs to 2D grids that CNNs can be directly applied to, while preserving graph topology as much as possible. We therefore propose two novel graph-to-grid mapping schemes, namely, {\em graph-preserving grid layout (GPGL)} and its extension {\em Hierarchical GPGL (H-GPGL)} for computational efficiency. We formulate the GPGL problem as integer programming and further propose an approximate yet efficient solver based on a penalized Kamada-Kawai method, a well-known optimization algorithm in 2D graph drawing. We propose a novel vertex separation penalty that encourages graph vertices to lay on the grid without any overlap. Along with this image representation, even extra 2D maxpooling layers contribute to the PointNet, a widely applied point-based neural network. We demonstrate the empirical success of GPGL on general graph classification with small graphs and H-GPGL on 3D point cloud segmentation with large graphs, based on 2D CNNs including VGG16, ResNet50 and multi-scale maxout (MSM) CNN.

31.Heuristic Weakly Supervised 3D Human Pose Estimation in Novel Contexts without Any 3D Pose Ground Truth ⬇️

Monocular 3D human pose estimation from a single RGB image has received a lot attentions in the past few year. Pose inference models with competitive performance however require supervision with 3D pose ground truth data or at least known pose priors in their target domain. Yet, these data requirements in many real-world applications with data collection constraints may not be achievable. In this paper, we present a heuristic weakly supervised solution, called HW-HuP to estimate 3D human pose in contexts that no ground truth 3D data is accessible, even for fine-tuning. HW-HuP learns partial pose priors from public 3D human pose datasets and uses easy-to-access observations from the target domain to iteratively estimate 3D human pose and shape in an optimization and regression hybrid cycle. In our design, depth data as an auxiliary information is employed as weak supervision during training, yet it is not needed for the inference. We evaluate HW-HuP performance qualitatively on datasets of both in-bed human and infant poses, where no ground truth 3D pose is provided neither any target prior. We also test HW-HuP performance quantitatively on a publicly available motion capture dataset against the 3D ground truth. HW-HuP is also able to be extended to other input modalities for pose estimation tasks especially under adverse vision conditions, such as occlusion or full darkness. On the Human3.6M benchmark, HW-HuP shows 104.1mm in MPJPE and 50.4mm in PA MPJPE, comparable to the existing state-of-the-art approaches that benefit from full 3D pose supervision.

32.Wisdom for the Crowd: Discoursive Power in Annotation Instructions for Computer Vision ⬇️

Developers of computer vision algorithms outsource some of the labor involved in annotating training data through business process outsourcing companies and crowdsourcing platforms. Many data annotators are situated in the Global South and are considered independent contractors. This paper focuses on the experiences of Argentinian and Venezuelan annotation workers. Through qualitative methods, we explore the discourses encoded in the task instructions that these workers follow to annotate computer vision datasets. Our preliminary findings indicate that annotation instructions reflect worldviews imposed on workers and, through their labor, on datasets. Moreover, we observe that for-profit goals drive task instructions and that managers and algorithms make sure annotations are done according to requesters' commands. This configuration presents a form of commodified labor that perpetuates power asymmetries while reinforcing social inequalities and is compelled to reproduce them into datasets and, subsequently, in computer vision systems.

33.Weakly Supervised Instance Attention for Multisource Fine-Grained Object Recognition ⬇️

Multisource image analysis that leverages complementary spectral, spatial, and structural information benefits fine-grained object recognition that aims to classify an object into one of many similar subcategories. However, for multisource tasks that involve relatively small objects, even the smallest registration errors can introduce high uncertainty in the classification process. We approach this problem from a weakly supervised learning perspective in which the input images correspond to larger neighborhoods around the expected object locations where an object with a given class label is present in the neighborhood without any knowledge of its exact location. The proposed method uses a single-source deep instance attention model with parallel branches for joint localization and classification of objects, and extends this model into a multisource setting where a reference source that is assumed to have no location uncertainty is used to aid the fusion of multiple sources in four different levels: probability level, logit level, feature level, and pixel level. We show that all levels of fusion provide higher accuracies compared to the state-of-the-art, with the best performing method of feature-level fusion resulting in 53% accuracy for the recognition of 40 different types of trees, corresponding to an improvement of 5.7% over the best performing baseline when RGB, multispectral, and LiDAR data are used. We also provide an in-depth comparison by evaluating each model at various parameter complexity settings, where the increased model capacity results in a further improvement of 6.3% over the default capacity setting.

34.HOME: Heatmap Output for future Motion Estimation ⬇️

In this paper, we propose HOME, a framework tackling the motion forecasting problem with an image output representing the probability distribution of the agent's future location. This method allows for a simple architecture with classic convolution networks coupled with attention mechanism for agent interactions, and outputs an unconstrained 2D top-view representation of the agent's possible future. Based on this output, we design two methods to sample a finite set of agent's future locations. These methods allow us to control the optimization trade-off between miss rate and final displacement error for multiple modalities without having to retrain any part of the model. We apply our method to the Argoverse Motion Forecasting Benchmark and achieve 1st place on the online leaderboard.

35.Boosting Crowd Counting with Transformers ⬇️

Significant progress on the crowd counting problem has been achieved by integrating larger context into convolutional neural networks (CNNs). This indicates that global scene context is essential, despite the seemingly bottom-up nature of the problem. This may be explained by the fact that context knowledge can adapt and improve local feature extraction to a given scene. In this paper, we therefore investigate the role of global context for crowd counting. Specifically, a pure transformer is used to extract features with global information from overlapping image patches. Inspired by classification, we add a context token to the input sequence, to facilitate information exchange with tokens corresponding to image patches throughout transformer layers. Due to the fact that transformers do not explicitly model the tried-and-true channel-wise interactions, we propose a token-attention module (TAM) to recalibrate encoded features through channel-wise attention informed by the context token. Beyond that, it is adopted to predict the total person count of the image through regression-token module (RTM). Extensive experiments demonstrate that our method achieves state-of-the-art performance on various datasets, including ShanghaiTech, UCF-QNRF, JHU-CROWD++ and NWPU. On the large-scale JHU-CROWD++ dataset, our method improves over the previous best results by 26.9% and 29.9% in terms of MAE and MSE, respectively.

36.COTR: Convolution in Transformer Network for End to End Polyp Detection ⬇️

Purpose: Colorectal cancer (CRC) is the second most common cause of cancer mortality worldwide. Colonoscopy is a widely used technique for colon screening and polyp lesions diagnosis. Nevertheless, manual screening using colonoscopy suffers from a substantial miss rate of polyps and is an overwhelming burden for endoscopists. Computer-aided diagnosis (CAD) for polyp detection has the potential to reduce human error and human burden. However, current polyp detection methods based on object detection framework need many handcrafted pre-processing and post-processing operations or user guidance that require domain-specific knowledge.
Methods: In this paper, we propose a convolution in transformer (COTR) network for end-to-end polyp detection. Motivated by the detection transformer (DETR), COTR is constituted by a CNN for feature extraction, transformer encoder layers interleaved with convolutional layers for feature encoding and recalibration, transformer decoder layers for object querying, and a feed-forward network for detection prediction. Considering the slow convergence of DETR, COTR embeds convolution layers into transformer encoder for feature reconstruction and convergence acceleration.
Results: Experimental results on two public polyp datasets show that COTR achieved 91.49% precision, 82.69% sensitivity, and 86.87% F1-score on the ETIS-LARIB, and 91.67% precision, 93.54% sensitivity, and 92.60% F1-score on the CVC-ColonDB.
Conclusion: This study proposed an end to end detection method based on detection transformer for colorectal polyp detection. Experimental results on ETIS-LARIB and CVC-ColonDB dataset demonstrated that the proposed model achieved comparable performance against state-of-the-art methods.

37.End-to-End Video Object Detection with Spatial-Temporal Transformers ⬇️

Recently, DETR and Deformable DETR have been proposed to eliminate the need for many hand-designed components in object detection while demonstrating good performance as previous complex hand-crafted detectors. However, their performance on Video Object Detection (VOD) has not been well explored. In this paper, we present TransVOD, an end-to-end video object detection model based on a spatial-temporal Transformer architecture. The goal of this paper is to streamline the pipeline of VOD, effectively removing the need for many hand-crafted components for feature aggregation, e.g., optical flow, recurrent neural networks, relation networks. Besides, benefited from the object query design in DETR, our method does not need complicated post-processing methods such as Seq-NMS or Tubelet rescoring, which keeps the pipeline simple and clean. In particular, we present temporal Transformer to aggregate both the spatial object queries and the feature memories of each frame. Our temporal Transformer consists of three components: Temporal Deformable Transformer Encoder (TDTE) to encode the multiple frame spatial details, Temporal Query Encoder (TQE) to fuse object queries, and Temporal Deformable Transformer Decoder to obtain current frame detection results. These designs boost the strong baseline deformable DETR by a significant margin (3%-4% mAP) on the ImageNet VID dataset. TransVOD yields comparable results performance on the benchmark of ImageNet VID. We hope our TransVOD can provide a new perspective for video object detection. Code will be made publicly available at this https URL.

38.Skeleton-aware multi-scale heatmap regression for 2D hand pose estimation ⬇️

Existing RGB-based 2D hand pose estimation methods learn the joint locations from a single resolution, which is not suitable for different hand sizes. To tackle this problem, we propose a new deep learning-based framework that consists of two main modules. The former presents a segmentation-based approach to detect the hand skeleton and localize the hand bounding box. The second module regresses the 2D joint locations through a multi-scale heatmap regression approach that exploits the predicted hand skeleton as a constraint to guide the model. Furthermore, we construct a new dataset that is suitable for both hand detection and pose estimation. We qualitatively and quantitatively validate our method on two datasets. Results demonstrate that the proposed method outperforms state-of-the-art and can recover the pose even in cluttered images and complex poses.

39.A hybrid classification-regression approach for 3D hand pose estimation using graph convolutional networks ⬇️

Hand pose estimation is a crucial part of a wide range of augmented reality and human-computer interaction applications. Predicting the 3D hand pose from a single RGB image is challenging due to occlusion and depth ambiguities. GCN-based (Graph Convolutional Networks) methods exploit the structural relationship similarity between graphs and hand joints to model kinematic dependencies between joints. These techniques use predefined or globally learned joint relationships, which may fail to capture pose-dependent constraints. To address this problem, we propose a two-stage GCN-based framework that learns per-pose relationship constraints. Specifically, the first phase quantizes the 2D/3D space to classify the joints into 2D/3D blocks based on their locality. This spatial dependency information guides this phase to estimate reliable 2D and 3D poses. The second stage further improves the 3D estimation through a GCN-based module that uses an adaptative nearest neighbor algorithm to determine joint relationships. Extensive experiments show that our multi-stage GCN approach yields an efficient model that produces accurate 2D/3D hand poses and outperforms the state-of-the-art on two public datasets.

40.VS-Net: Voting with Segmentation for Visual Localization ⬇️

Visual localization is of great importance in robotics and computer vision. Recently, scene coordinate regression based methods have shown good performance in visual localization in small static scenes. However, it still estimates camera poses from many inferior scene coordinates. To address this problem, we propose a novel visual localization framework that establishes 2D-to-3D correspondences between the query image and the 3D map with a series of learnable scene-specific landmarks. In the landmark generation stage, the 3D surfaces of the target scene are over-segmented into mosaic patches whose centers are regarded as the scene-specific landmarks. To robustly and accurately recover the scene-specific landmarks, we propose the Voting with Segmentation Network (VS-Net) to segment the pixels into different landmark patches with a segmentation branch and estimate the landmark locations within each patch with a landmark location voting branch. Since the number of landmarks in a scene may reach up to 5000, training a segmentation network with such a large number of classes is both computation and memory costly for the commonly used cross-entropy loss. We propose a novel prototype-based triplet loss with hard negative mining, which is able to train semantic segmentation networks with a large number of labels efficiently. Our proposed VS-Net is extensively tested on multiple public benchmarks and can outperform state-of-the-art visual localization methods. Code and models are available at \href{this https URL}{this https URL}.

41.Weakly-supervised Cross-view 3D Human Pose Estimation ⬇️

Although monocular 3D human pose estimation methods have made significant progress, it's far from being solved due to the inherent depth ambiguity. Instead, exploiting multi-view information is a practical way to achieve absolute 3D human pose estimation. In this paper, we propose a simple yet effective pipeline for weakly-supervised cross-view 3D human pose estimation. By only using two camera views, our method can achieve state-of-the-art performance in a weakly-supervised manner, requiring no 3D ground truth but only 2D annotations. Specifically, our method contains two steps: triangulation and refinement. First, given the 2D keypoints that can be obtained through any classic 2D detection methods, triangulation is performed across two views to lift the 2D keypoints into coarse 3D poses.Then, a novel cross-view U-shaped graph convolutional network (CV-UGCN), which can explore the spatial configurations and cross-view correlations, is designed to refine the coarse 3D poses. In particular, the refinement progress is achieved through weakly-supervised learning, in which geometric and structure-aware consistency checks are performed. We evaluate our method on the standard benchmark dataset, Human3.6M. The Mean Per Joint Position Error on the benchmark dataset is 27.4 mm, which outperforms the state-of-the-arts remarkably (27.4 mm vs 30.2 mm).

42.CMUA-Watermark: A Cross-Model Universal Adversarial Watermark for Combating Deepfakes ⬇️

Malicious application of deepfakes (i.e., technologies can generate target faces or face attributes) has posed a huge threat to our society. The fake multimedia content generated by deepfake models can harm the reputation and even threaten the property of the person who has been impersonated. Fortunately, the adversarial watermark could be used for combating deepfake models, leading them to generate distorted images. The existing methods require an individual training process for every facial image, to generate the adversarial watermark against a specific deepfake model, which are extremely inefficient. To address this problem, we propose a universal adversarial attack method on deepfake models, to generate a Cross-Model Universal Adversarial Watermark (CMUA-Watermark) that can protect thousands of facial images from multiple deepfake models. Specifically, we first propose a cross-model universal attack pipeline by attacking multiple deepfake models and combining gradients from these models iteratively. Then we introduce a batch-based method to alleviate the conflict of adversarial watermarks generated by different facial images. Finally, we design a more reasonable and comprehensive evaluation method for evaluating the effectiveness of the adversarial watermark. Experimental results demonstrate that the proposed CMUA-Watermark can effectively distort the fake facial images generated by deepfake models and successfully protect facial images from deepfakes in real scenes.

43.FCCDN: Feature Constraint Network for VHR Image Change Detection ⬇️

Change detection is the process of identifying pixel-wise differences of bi-temporal co-registered images. It is of great significance to Earth observation. Recently, with the emerging of deep learning (DL), deep convolutional neural networks (CNNs) based methods have shown their power and feasibility in the field of change detection. However, there is still a lack of effective supervision for change feature learning. In this work, a feature constraint change detection network (FCCDN) is proposed. We constrain features both on bi-temporal feature extraction and feature fusion. More specifically, we propose a dual encoder-decoder network backbone for the change detection task. At the center of the backbone, we design a non-local feature pyramid network to extract and fuse multi-scale features. To fuse bi-temporal features in a robust way, we build a dense connection-based feature fusion module. Moreover, a self-supervised learning-based strategy is proposed to constrain feature learning. Based on FCCDN, we achieve state-of-the-art performance on two building change detection datasets (LEVIR-CD and WHU). On the LEVIR-CD dataset, we achieve IoU of 0.8569 and F1 score of 0.9229. On the WHU dataset, we achieve IoU of 0.8820 and F1 score of 0.9373. Moreover, we, for the first time, achieve the acquire of accurate bi-temporal semantic segmentation results without using semantic segmentation labels. It is vital for the application of change detection because it saves the cost of labeling.

44.Coarse to Fine Multi-Resolution Temporal Convolutional Network ⬇️

Temporal convolutional networks (TCNs) are a commonly used architecture for temporal video segmentation. TCNs however, tend to suffer from over-segmentation errors and require additional refinement modules to ensure smoothness and temporal coherency. In this work, we propose a novel temporal encoder-decoder to tackle the problem of sequence fragmentation. In particular, the decoder follows a coarse-to-fine structure with an implicit ensemble of multiple temporal resolutions. The ensembling produces smoother segmentations that are more accurate and better-calibrated, bypassing the need for additional refinement modules. In addition, we enhance our training with a multi-resolution feature-augmentation strategy to promote robustness to varying temporal resolutions. Finally, to support our architecture and encourage further sequence coherency, we propose an action loss that penalizes misclassifications at the video level. Experiments show that our stand-alone architecture, together with our novel feature-augmentation strategy and new loss, outperforms the state-of-the-art on three temporal video segmentation benchmarks.

45.Exploring Robustness of Unsupervised Domain Adaptation in Semantic Segmentation ⬇️

Recent studies imply that deep neural networks are vulnerable to adversarial examples -- inputs with a slight but intentional perturbation are incorrectly classified by the network. Such vulnerability makes it risky for some security-related applications (e.g., semantic segmentation in autonomous cars) and triggers tremendous concerns on the model reliability. For the first time, we comprehensively evaluate the robustness of existing UDA methods and propose a robust UDA approach. It is rooted in two observations: (i) the robustness of UDA methods in semantic segmentation remains unexplored, which pose a security concern in this field; and (ii) although commonly used self-supervision (e.g., rotation and jigsaw) benefits image tasks such as classification and recognition, they fail to provide the critical supervision signals that could learn discriminative representation for segmentation tasks. These observations motivate us to propose adversarial self-supervision UDA (or ASSUDA) that maximizes the agreement between clean images and their adversarial examples by a contrastive loss in the output space. Extensive empirical studies on commonly used benchmarks demonstrate that ASSUDA is resistant to adversarial attacks.

46.Adapted Human Pose: Monocular 3D Human Pose Estimation with Zero Real 3D Pose Data ⬇️

The ultimate goal for an inference model is to be robust and functional in real life applications. However, training vs. test data domain gaps often negatively affect model performance. This issue is especially critical for the monocular 3D human pose estimation problem, in which 3D human data is often collected in a controlled lab setting. In this paper, we focus on alleviating the negative effect of domain shift by presenting our adapted human pose (AHuP) approach that addresses adaptation problems in both appearance and pose spaces. AHuP is built around a practical assumption that in real applications, data from target domain could be inaccessible or only limited information can be acquired. We illustrate the 3D pose estimation performance of AHuP in two scenarios. First, when source and target data differ significantly in both appearance and pose spaces, in which we learn from synthetic 3D human data (with zero real 3D human data) and show comparable performance with the state-of-the-art 3D pose estimation models that have full access to the real 3D human pose benchmarks for training. Second, when source and target datasets differ mainly in the pose space, in which AHuP approach can be applied to further improve the performance of the state-of-the-art models when tested on the datasets different from their training dataset.

47.Stereo Matching Based on Visual Sensitive Information ⬇️

The area of computer vision is one of the most discussed topics amongst many scholars, and stereo matching is its most important sub fields. After the parallax map is transformed into a depth map, it can be applied to many intelligent fields. In this paper, a stereo matching algorithm based on visual sensitive information is proposed by using standard images from Middlebury dataset. Aiming at the limitation of traditional stereo matching algorithms regarding the cost window, a cost aggregation algorithm based on the dynamic window is proposed, and the disparity image is optimized by using left and right consistency detection to further reduce the error matching rate. The experimental results show that the proposed algorithm can effectively enhance the stereo matching effect of the image providing significant improvement in accuracy as compared with the classical census algorithm. The proposed model code, dataset, and experimental results are available at this https URL.

48.Texture synthesis via projection onto multiscale, multilayer statistics ⬇️

We provide a new model for texture synthesis based on a multiscale, multilayer feature extractor. Within the model, textures are represented by a set of statistics computed from ReLU wavelet coefficients at different layers, scales and orientations. A new image is synthesized by matching the target statistics via an iterative projection algorithm. We explain the necessity of the different types of pre-defined wavelet filters used in our model and the advantages of multilayer structures for image synthesis. We demonstrate the power of our model by generating samples of high quality textures and providing insights into deep representations for texture images.

49.GOO: A Dataset for Gaze Object Prediction in Retail Environments ⬇️

One of the most fundamental and information-laden actions humans do is to look at objects. However, a survey of current works reveals that existing gaze-related datasets annotate only the pixel being looked at, and not the boundaries of a specific object of interest. This lack of object annotation presents an opportunity for further advancing gaze estimation research. To this end, we present a challenging new task called gaze object prediction, where the goal is to predict a bounding box for a person's gazed-at object. To train and evaluate gaze networks on this task, we present the Gaze On Objects (GOO) dataset. GOO is composed of a large set of synthetic images (GOO Synth) supplemented by a smaller subset of real images (GOO-Real) of people looking at objects in a retail environment. Our work establishes extensive baselines on GOO by re-implementing and evaluating selected state-of-the art models on the task of gaze following and domain adaptation. Code is available on github.

50.PLM: Partial Label Masking for Imbalanced Multi-label Classification ⬇️

Neural networks trained on real-world datasets with long-tailed label distributions are biased towards frequent classes and perform poorly on infrequent classes. The imbalance in the ratio of positive and negative samples for each class skews network output probabilities further from ground-truth distributions. We propose a method, Partial Label Masking (PLM), which utilizes this ratio during training. By stochastically masking labels during loss computation, the method balances this ratio for each class, leading to improved recall on minority classes and improved precision on frequent classes. The ratio is estimated adaptively based on the network's performance by minimizing the KL divergence between predicted and ground-truth distributions. Whereas most existing approaches addressing data imbalance are mainly focused on single-label classification and do not generalize well to the multi-label case, this work proposes a general approach to solve the long-tail data imbalance issue for multi-label classification. PLM is versatile: it can be applied to most objective functions and it can be used alongside other strategies for class imbalance. Our method achieves strong performance when compared to existing methods on both multi-label (MultiMNIST and MSCOCO) and single-label (imbalanced CIFAR-10 and CIFAR-100) image classification datasets.

51.PAL: Intelligence Augmentation using Egocentric Visual Context Detection ⬇️

Egocentric visual context detection can support intelligence augmentation applications. We created a wearable system, called PAL, for wearable, personalized, and privacy-preserving egocentric visual context detection. PAL has a wearable device with a camera, heart-rate sensor, on-device deep learning, and audio input/output. PAL also has a mobile/web application for personalized context labeling. We used on-device deep learning models for generic object and face detection, low-shot custom face and context recognition (e.g., activities like brushing teeth), and custom context clustering (e.g., indoor locations). The models had over 80% accuracy in in-the-wild contexts (~1000 images) and we tested PAL for intelligence augmentation applications like behavior change. We have made PAL is open-source to further support intelligence augmentation using personalized and privacy-preserving egocentric visual contexts.

52.Automated Knee X-ray Report Generation ⬇️

Gathering manually annotated images for the purpose of training a predictive model is far more challenging in the medical domain than for natural images as it requires the expertise of qualified radiologists. We therefore propose to take advantage of past radiological exams (specifically, knee X-ray examinations) and formulate a framework capable of learning the correspondence between the images and reports, and hence be capable of generating diagnostic reports for a given X-ray examination consisting of an arbitrary number of image views. We demonstrate how aggregating the image features of individual exams and using them as conditional inputs when training a language generation model results in auto-generated exam reports that correlate well with radiologist-generated reports.

53.Soccer Player Tracking in Low Quality Video ⬇️

In this paper we propose a system capable of tracking multiple soccer players in different types of video quality. The main goal, in contrast to most state-of-art soccer player tracking systems, is the ability of execute effectively tracking in videos of low-quality. We adapted a state-of-art Multiple Object Tracking to the task. In order to do that adaptation, we created a Detection and a Tracking Dataset for 3 different qualities of video. The results of our system are conclusive of its high performance.

54.ADNet: Attention-guided Deformable Convolutional Network for High Dynamic Range Imaging ⬇️

In this paper, we present an attention-guided deformable convolutional network for hand-held multi-frame high dynamic range (HDR) imaging, namely ADNet. This problem comprises two intractable challenges of how to handle saturation and noise properly and how to tackle misalignments caused by object motion or camera jittering. To address the former, we adopt a spatial attention module to adaptively select the most appropriate regions of various exposure low dynamic range (LDR) images for fusion. For the latter one, we propose to align the gamma-corrected images in the feature-level with a Pyramid, Cascading and Deformable (PCD) alignment module. The proposed ADNet shows state-of-the-art performance compared with previous methods, achieving a PSNR-$l$ of 39.4471 and a PSNR-$\mu$ of 37.6359 in NTIRE 2021 Multi-Frame HDR Challenge.

55.Towards Automatic Recognition of Pure & Mixed Stones using Intraoperative Endoscopic Digital Images ⬇️

Objective: To assess automatic computer-aided in-situ recognition of morphological features of pure and mixed urinary stones using intraoperative digital endoscopic images acquired in a clinical setting. Materials and methods: In this single-centre study, an experienced urologist intraoperatively and prospectively examined the surface and section of all kidney stones encountered. Calcium oxalate monohydrate (COM/Ia), dihydrate (COD/IIb) and uric acid (UA/IIIb) morphological criteria were collected and classified to generate annotated datasets. A deep convolutional neural network (CNN) was trained to predict the composition of both pure and mixed stones. To explain the predictions of the deep neural network model, coarse localisation heat-maps were plotted to pinpoint key areas identified by the network. Results: This study included 347 and 236 observations of stone surface and stone section, respectively. A highest sensitivity of 98 % was obtained for the type "pure IIIb/UA" using surface images. The most frequently encountered morphology was that of the type "pure Ia/COM"; it was correctly predicted in 91 % and 94 % of cases using surface and section images, respectively. Of the mixed type "Ia/COM+IIb/COD", Ia/COM was predicted in 84 % of cases using surface images, IIb/COD in 70 % of cases, and both in 65 % of cases. Concerning mixed Ia/COM+IIIb/UA stones, Ia/COM was predicted in 91 % of cases using section images, IIIb/UA in 69 % of cases, and both in 74 % of cases. Conclusions: This preliminary study demonstrates that deep convolutional neural networks are promising to identify kidney stone composition from endoscopic images acquired intraoperatively. Both pure and mixed stone composition could be discriminated. Collected in a clinical setting, surface and section images analysed by deep CNN provide valuable information about stone morphology for computer-aided diagnosis.

56.Video-based Person Re-identification without Bells and Whistles ⬇️

Video-based person re-identification (Re-ID) aims at matching the video tracklets with cropped video frames for identifying the pedestrians under different cameras. However, there exists severe spatial and temporal misalignment for those cropped tracklets due to the imperfect detection and tracking results generated with obsolete methods. To address this issue, we present a simple re-Detect and Link (DL) module which can effectively reduce those unexpected noise through applying the deep learning-based detection and tracking on the cropped tracklets. Furthermore, we introduce an improved model called Coarse-to-Fine Axial-Attention Network (CF-AAN). Based on the typical Non-local Network, we replace the non-local module with three 1-D position-sensitive axial attentions, in addition to our proposed coarse-to-fine structure. With the developed CF-AAN, compared to the original non-local operation, we can not only significantly reduce the computation cost but also obtain the state-of-the-art performance (91.3% in rank-1 and 86.5% in mAP) on the large-scale MARS dataset. Meanwhile, by simply adopting our DL module for data alignment, to our surprise, several baseline models can achieve better or comparable results with the current state-of-the-arts. Besides, we discover the errors not only for the identity labels of tracklets but also for the evaluation protocol for the test data of MARS. We hope that our work can help the community for the further development of invariant representation without the hassle of the spatial and temporal alignment and dataset noise. The code, corrected labels, evaluation protocol, and the aligned data will be available at this https URL.

57.Semi-Supervised Few-Shot Classification with Deep Invertible Hybrid Models ⬇️

In this paper, we propose a deep invertible hybrid model which integrates discriminative and generative learning at a latent space level for semi-supervised few-shot classification. Various tasks for classifying new species from image data can be modeled as a semi-supervised few-shot classification, which assumes a labeled and unlabeled training examples and a small support set of the target classes. Predicting target classes with a few support examples per class makes the learning task difficult for existing semi-supervised classification methods, including selftraining, which iteratively estimates class labels of unlabeled training examples to learn a classifier for the training classes. To exploit unlabeled training examples effectively, we adopt as the objective function the composite likelihood, which integrates discriminative and generative learning and suits better with deep neural networks than the parameter coupling prior, the other popular integrated learning approach. In our proposed model, the discriminative and generative models are respectively Prototypical Networks, which have shown excellent performance in various kinds of few-shot learning, and Normalizing Flow a deep invertible model which returns the exact marginal likelihood unlike the other three major methods, i.e., VAE, GAN, and autoregressive model. Our main originality lies in our integration of these components at a latent space level, which is effective in preventing overfitting. Experiments using mini-ImageNet and VGG-Face datasets show that our method outperforms selftraining based Prototypical Networks.

58.Revisiting Knowledge Distillation for Object Detection ⬇️

The existing solutions for object detection distillation rely on the availability of both a teacher model and ground-truth labels. We propose a new perspective to relax this constraint. In our framework, a student is first trained with pseudo labels generated by the teacher, and then fine-tuned using labeled data, if any available. Extensive experiments demonstrate improvements over existing object detection distillation algorithms. In addition, decoupling the teacher and ground-truth distillation in this framework provides interesting properties such: as 1) using unlabeled data to further improve the student's performance, 2) combining multiple teacher models of different architectures, even with different object categories, and 3) reducing the need for labeled data (with only 20% of COCO labels, this method achieves the same performance as the model trained on the entire set of labels). Furthermore, a by-product of this approach is the potential usage for domain adaptation. We verify these properties through extensive experiments.

59.Searching Collaborative Agents for Multi-plane Localization in 3D Ultrasound ⬇️

3D ultrasound (US) has become prevalent due to its rich spatial and diagnostic information not contained in 2D US. Moreover, 3D US can contain multiple standard planes (SPs) in one shot. Thus, automatically localizing SPs in 3D US has the potential to improve user-independence and scanning-efficiency. However, manual SP localization in 3D US is challenging because of the low image quality, huge search space and large anatomical variability. In this work, we propose a novel multi-agent reinforcement learning (MARL) framework to simultaneously localize multiple SPs in 3D US. Our contribution is four-fold. First, our proposed method is general and it can accurately localize multiple SPs in different challenging US datasets. Second, we equip the MARL system with a recurrent neural network (RNN) based collaborative module, which can strengthen the communication among agents and learn the spatial relationship among planes effectively. Third, we explore to adopt the neural architecture search (NAS) to automatically design the network architecture of both the agents and the collaborative module. Last, we believe we are the first to realize automatic SP localization in pelvic US volumes, and note that our approach can handle both normal and abnormal uterus cases. Extensively validated on two challenging datasets of the uterus and fetal brain, our proposed method achieves the average localization accuracy of 7.03 degrees/1.59mm and 9.75 degrees/1.19mm. Experimental results show that our light-weight MARL model has higher accuracy than state-of-the-art methods.

60.HPNet: Deep Primitive Segmentation Using Hybrid Representations ⬇️

This paper introduces HPNet, a novel deep-learning approach for segmenting a 3D shape represented as a point cloud into primitive patches. The key to deep primitive segmentation is learning a feature representation that can separate points of different primitives. Unlike utilizing a single feature representation, HPNet leverages hybrid representations that combine one learned semantic descriptor, two spectral descriptors derived from predicted geometric parameters, as well as an adjacency matrix that encodes sharp edges. Moreover, instead of merely concatenating the descriptors, HPNet optimally combines hybrid representations by learning combination weights. This weighting module builds on the entropy of input features. The output primitive segmentation is obtained from a mean-shift clustering module. Experimental results on benchmark datasets ANSI and ABCParts show that HPNet leads to significant performance gains from baseline approaches.

61.Embracing New Techniques in Deep Learning for Estimating Image Memorability ⬇️

Various work has suggested that the memorability of an image is consistent across people, and thus can be treated as an intrinsic property of an image. Using computer vision models, we can make specific predictions about what people will remember or forget. While older work has used now-outdated deep learning architectures to predict image memorability, innovations in the field have given us new techniques to apply to this problem. Here, we propose and evaluate five alternative deep learning models which exploit developments in the field from the last five years, largely the introduction of residual neural networks, which are intended to allow the model to use semantic information in the memorability estimation process. These new models were tested against the prior state of the art with a combined dataset built to optimize both within-category and across-category predictions. Our findings suggest that the key prior memorability network had overstated its generalizability and was overfit on its training set. Our new models outperform this prior model, leading us to conclude that Residual Networks outperform simpler convolutional neural networks in memorability regression. We make our new state-of-the-art model readily available to the research community, allowing memory researchers to make predictions about memorability on a wider range of images.

62.Puck localization and multi-task event recognition in broadcast hockey videos ⬇️

Puck localization is an important problem in ice hockey video analytics useful for analyzing the game, determining play location, and assessing puck possession. The problem is challenging due to the small size of the puck, excessive motion blur due to high puck velocity and occlusions due to players and boards. In this paper, we introduce and implement a network for puck localization in broadcast hockey video. The network leverages expert NHL play-by-play annotations and uses temporal context to locate the puck. Player locations are incorporated into the network through an attention mechanism by encoding player positions with a Gaussian-based spatial heatmap drawn at player positions. Since event occurrence on the rink and puck location are related, we also perform event recognition by augmenting the puck localization network with an event recognition head and training the network through multi-task learning. Experimental results demonstrate that the network is able to localize the puck with an AUC of $73.1 %$ on the test set. The puck location can be inferred in 720p broadcast videos at $5$ frames per second. It is also demonstrated that multi-task learning with puck location improves event recognition accuracy.

63.BCNet: Searching for Network Width with Bilaterally Coupled Network ⬇️

Searching for a more compact network width recently serves as an effective way of channel pruning for the deployment of convolutional neural networks (CNNs) under hardware constraints. To fulfill the searching, a one-shot supernet is usually leveraged to efficiently evaluate the performance \wrt~different network widths. However, current methods mainly follow a \textit{unilaterally augmented} (UA) principle for the evaluation of each width, which induces the training unfairness of channels in supernet. In this paper, we introduce a new supernet called Bilaterally Coupled Network (BCNet) to address this issue. In BCNet, each channel is fairly trained and responsible for the same amount of network widths, thus each network width can be evaluated more accurately. Besides, we leverage a stochastic complementary strategy for training the BCNet, and propose a prior initial population sampling method to boost the performance of the evolutionary search. Extensive experiments on benchmark CIFAR-10 and ImageNet datasets indicate that our method can achieve state-of-the-art or competing performance over other baseline methods. Moreover, our method turns out to further boost the performance of NAS models by refining their network widths. For example, with the same FLOPs budget, our obtained EfficientNet-B0 achieves 77.36% Top-1 accuracy on ImageNet dataset, surpassing the performance of original setting by 0.48%.

64.SmartPatch: Improving Handwritten Word Imitation with Patch Discriminators ⬇️

As of recent generative adversarial networks have allowed for big leaps in the realism of generated images in diverse domains, not the least of which being handwritten text generation. The generation of realistic-looking hand-written text is important because it can be used for data augmentation in handwritten text recognition (HTR) systems or human-computer interaction. We propose SmartPatch, a new technique increasing the performance of current state-of-the-art methods by augmenting the training feedback with a tailored solution to mitigate pen-level artifacts. We combine the well-known patch loss with information gathered from the parallel trained handwritten text recognition system and the separate characters of the word. This leads to a more enhanced local discriminator and results in more realistic and higher-quality generated handwritten words.

65.Design to automate the detection and counting of Tuberculosis(TB) bacilli ⬇️

Tuberculosis is a contagious disease which is one of the leading causes of death, globally. The general diagnosis methods for tuberculosis include microscopic examination, tuberculin skin test, culture method, enzyme linked immunosorbent assay (ELISA) and electronic nose system. World Health Organization (WHO) recommends standard microscopic examination for early diagnosis of tuberculosis. In microscopy, the technician examines field of views (FOVs) in sputum smear for presence of any TB bacilli and counts the number of TB bacilli per FOV to report the level of severity. This process is time consuming with an increased concentration for an experienced staff to examine a single sputum smear. The examination demands for skilled technicians in high-prevalence countries which may lead to overload, fatigue and diminishes the quality of microscopy. Thus, a computer assisted system is proposed and designed for the detection of tuberculosis bacilli to assist pathologists with increased sensitivity and specificity. The manual efforts in detecting and counting the number of TB bacilli is greatly minimized. The system obtains Ziehl-Neelsen stained microscopic images from conventional microscope at 100x magnification and passes the data to the detection system. Initially the segmentation of TB bacilli was done using RGB thresholding and Sauvola's adaptive thresholding algorithm. To eliminate the non-TB bacilli from coarse level segmentation, shape descriptors like area, perimeter, convex hull, major axis length and eccentricity are used to extract only the TB bacilli features. Finally, the TB bacilli are counted using the generated bounding boxes to report the level of severity.

66.DDR-Net: Dividing and Downsampling Mixed Network for Diffeomorphic Image Registration ⬇️

Deep diffeomorphic registration faces significant challenges for high-dimensional images, especially in terms of memory limits. Existing approaches either downsample original images, or approximate underlying transformations, or reduce model size. The information loss during the approximation or insufficient model capacity is a hindrance to the registration accuracy for high-dimensional images, e.g., 3D medical volumes. In this paper, we propose a Dividing and Downsampling mixed Registration network (DDR-Net), a general architecture that preserves most of the image information at multiple scales. DDR-Net leverages the global context via downsampling the input and utilizes the local details from divided chunks of the input images. This design reduces the network input size and its memory cost; meanwhile, by fusing global and local information, DDR-Net obtains both coarse-level and fine-level alignments in the final deformation fields. We evaluate DDR-Net on three public datasets, i.e., OASIS, IBSR18, and 3DIRCADB-01, and the experimental results demonstrate our approach outperforms existing approaches.

67.Brain tumour segmentation using a triplanar ensemble of U-Nets ⬇️

Gliomas appear with wide variation in their characteristics both in terms of their appearance and location on brain MR images, which makes robust tumour segmentation highly challenging, and leads to high inter-rater variability even in manual segmentations. In this work, we propose a triplanar ensemble network, with an independent tumour core prediction module, for accurate segmentation of these tumours and their sub-regions. On evaluating our method on the MICCAI Brain Tumor Segmentation (BraTS) challenge validation dataset, for tumour sub-regions, we achieved a Dice similarity coefficient of 0.77 for both enhancing tumour (ET) and tumour core (TC). In the case of the whole tumour (WT) region, we achieved a Dice value of 0.89, which is on par with the top-ranking methods from BraTS'17-19. Our method achieved an evaluation score that was the equal 5th highest value (with our method ranking in 10th place) in the BraTS'20 challenge, with mean Dice values of 0.81, 0.89 and 0.84 on ET, WT and TC regions respectively on the BraTS'20 unseen test dataset.

68.Coarse-to-Fine for Sim-to-Real: Sub-Millimetre Precision Across the Workspace ⬇️

When training control policies for robot manipulation via deep learning, sim-to-real transfer can help satisfy the large data requirements. In this paper, we study the problem of zero-shot sim-to-real when the task requires both highly precise control, with sub-millimetre error tolerance, and full workspace generalisation. Our framework involves a coarse-to-fine controller, where trajectories initially begin with classical motion planning based on pose estimation, and transition to an end-to-end controller which maps images to actions and is trained in simulation with domain randomisation. In this way, we achieve precise control whilst also generalising the controller across the workspace and keeping the generality and robustness of vision-based, end-to-end control. Real-world experiments on a range of different tasks show that, by exploiting the best of both worlds, our framework significantly outperforms purely motion planning methods, and purely learning-based methods. Furthermore, we answer a range of questions on best practices for precise sim-to-real transfer, such as how different image sensor modalities and image feature representations perform.

69.Generation of COVID-19 Chest CT Scan Images using Generative Adversarial Networks ⬇️

SARS-CoV-2, also known as COVID-19 or Coronavirus, is a viral contagious disease that is infected by a novel coronavirus, and has been rapidly spreading across the globe. It is very important to test and isolate people to reduce spread, and from here comes the need to do this quickly and efficiently. According to some studies, Chest-CT outperforms RT-PCR lab testing, which is the current standard, when diagnosing COVID-19 patients. Due to this, computer vision researchers have developed various deep learning systems that can predict COVID-19 using a Chest-CT scan correctly to a certain degree. The accuracy of these systems is limited since deep learning neural networks such as CNNs (Convolutional Neural Networks) need a significantly large quantity of data for training in order to produce good quality results. Since the disease is relatively recent and more focus has been on CXR (Chest XRay) images, the available chest CT Scan image dataset is much less. We propose a method, by utilizing GANs, to generate synthetic chest CT images of both positive and negative COVID-19 patients. Using a pre-built predictive model, we concluded that around 40% of the generated images are correctly predicted as COVID-19 positive. The dataset thus generated can be used to train a CNN-based classifier which can help determine COVID-19 in a patient with greater accuracy.

70.A self-supervised learning strategy for postoperative brain cavity segmentation simulating resections ⬇️

Accurate segmentation of brain resection cavities (RCs) aids in postoperative analysis and determining follow-up treatment. Convolutional neural networks (CNNs) are the state-of-the-art image segmentation technique, but require large annotated datasets for training. Annotation of 3D medical images is time-consuming, requires highly-trained raters, and may suffer from high inter-rater variability. Self-supervised learning strategies can leverage unlabeled data for training.
We developed an algorithm to simulate resections from preoperative magnetic resonance images (MRIs). We performed self-supervised training of a 3D CNN for RC segmentation using our simulation method. We curated EPISURG, a dataset comprising 430 postoperative and 268 preoperative MRIs from 430 refractory epilepsy patients who underwent resective neurosurgery. We fine-tuned our model on three small annotated datasets from different institutions and on the annotated images in EPISURG, comprising 20, 33, 19 and 133 subjects.
The model trained on data with simulated resections obtained median (interquartile range) Dice score coefficients (DSCs) of 81.7 (16.4), 82.4 (36.4), 74.9 (24.2) and 80.5 (18.7) for each of the four datasets. After fine-tuning, DSCs were 89.2 (13.3), 84.1 (19.8), 80.2 (20.1) and 85.2 (10.8). For comparison, inter-rater agreement between human annotators from our previous study was 84.0 (9.9).
We present a self-supervised learning strategy for 3D CNNs using simulated RCs to accurately segment real RCs on postoperative MRI. Our method generalizes well to data from different institutions, pathologies and modalities. Source code, segmentation models and the EPISURG dataset are available at this https URL .

71.Pulmonary embolism identification in computerized tomography pulmonary angiography scans with deep learning technologies in COVID-19 patients ⬇️

The main objective of this work is to utilize state-of-the-art deep learning approaches for the identification of pulmonary embolism in CTPA-Scans for COVID-19 patients, provide an initial assessment of their performance and, ultimately, provide a fast-track prototype solution (system). We adopted and assessed some of the most popular convolutional neural network architectures through transfer learning approaches, to strive to combine good model accuracy with fast training. Additionally, we exploited one of the most popular one-stage object detection models for the localization (through object detection) of the pulmonary embolism regions-of-interests. The models of both approaches are trained on an original CTPA-Scan dataset, where we annotated of 673 CTPA-Scan images with 1,465 bounding boxes in total, highlighting pulmonary embolism regions-of-interests. We provide a brief assessment of some state-of-the-art image classification models by achieving validation accuracies of 91% in pulmonary embolism classification. Additionally, we achieved a precision of about 68% on average in the object detection model for the pulmonary embolism localization under 50% IoU threshold. For both approaches, we provide the entire training pipelines for future studies (step by step processes through source code). In this study, we present some of the most accurate and fast deep learning models for pulmonary embolism identification in CTPA-Scans images, through classification and localization (object detection) approaches for patients infected by COVID-19. We provide a fast-track solution (system) for the research community of the area, which combines both classification and object detection models for improving the precision of identifying pulmonary embolisms.

72.Smart mobile microscopy: towards fully-automated digitization ⬇️

Mobile microscopy is a newly formed field that emerged from a combination of optical microscopy capabilities and spread, functionality, and ever-increasing computing resources of mobile devices. Despite the idea of creating a system that would successfully merge a microscope, numerous computer vision methods, and a mobile device is regularly examined, the resulting implementations still require the presence of a qualified operator to control specimen digitization. In this paper, we address the task of surpassing this constraint and present a ``smart'' mobile microscope concept aimed at automatic digitization of the most valuable visual information about the specimen. We perform this through combining automated microscope setup control and classic techniques such as auto-focusing, in-focus filtering, and focus-stacking -- adapted and optimized as parts of a mobile cross-platform library.

73.AirNet: Neural Network Transmission over the Air ⬇️

State-of-the-art performance for many emerging edge applications is achieved by deep neural networks (DNNs). Often, these DNNs are location and time sensitive, and the parameters of a specific DNN must be delivered from an edge server to the edge device rapidly and efficiently to carry out time-sensitive inference tasks. We introduce AirNet, a novel training and analog transmission method that allows efficient wireless delivery of DNNs. We first train the DNN with noise injection to counter the wireless channel noise. We also employ pruning to reduce the channel bandwidth necessary for transmission, and perform knowledge distillation from a larger model to achieve satisfactory performance, despite the channel perturbations. We show that AirNet achieves significantly higher test accuracy compared to digital alternatives under the same bandwidth and power constraints. It also exhibits graceful degradation with channel quality, which reduces the requirement for accurate channel estimation.

74.Post-Training Sparsity-Aware Quantization ⬇️

Quantization is a technique used in deep neural networks (DNNs) to increase execution performance and hardware efficiency. Uniform post-training quantization (PTQ) methods are common, since they can be implemented efficiently in hardware and do not require extensive hardware resources or a training set. Mapping FP32 models to INT8 using uniform PTQ yields models with negligible accuracy degradation; however, reducing precision below 8 bits with PTQ is challenging, as accuracy degradation becomes noticeable, due to the increase in quantization noise. In this paper, we propose a sparsity-aware quantization (SPARQ) method, in which the unstructured and dynamic activation sparsity is leveraged in different representation granularities. 4-bit quantization, for example, is employed by dynamically examining the bits of 8-bit values and choosing a window of 4 bits, while first skipping zero-value bits. Moreover, instead of quantizing activation-by-activation to 4 bits, we focus on pairs of 8-bit activations and examine whether one of the two is equal to zero. If one is equal to zero, the second can opportunistically use the other's 4-bit budget; if both do not equal zero, then each is dynamically quantized to 4 bits, as described. SPARQ achieves minor accuracy degradation, 2x speedup over widely used hardware architectures, and a practical hardware implementation. The code is available at this https URL.

75.FBI-Denoiser: Fast Blind Image Denoiser for Poisson-Gaussian Noise ⬇️

We consider the challenging blind denoising problem for Poisson-Gaussian noise, in which no additional information about clean images or noise level parameters is available. Particularly, when only "single" noisy images are available for training a denoiser, the denoising performance of existing methods was not satisfactory. Recently, the blind pixelwise affine image denoiser (BP-AIDE) was proposed and significantly improved the performance in the above setting, to the extent that it is competitive with denoisers which utilized additional information. However, BP-AIDE seriously suffered from slow inference time due to the inefficiency of noise level estimation procedure and that of the blind-spot network (BSN) architecture it used. To that end, we propose Fast Blind Image Denoiser (FBI-Denoiser) for Poisson-Gaussian noise, which consists of two neural network models; 1) PGE-Net that estimates Poisson-Gaussian noise parameters 2000 times faster than the conventional methods and 2) FBI-Net that realizes a much more efficient BSN for pixelwise affine denoiser in terms of the number of parameters and inference speed. Consequently, we show that our FBI-Denoiser blindly trained solely based on single noisy images can achieve the state-of-the-art performance on several real-world noisy image benchmark datasets with much faster inference time (x 10), compared to BP-AIDE. The official code of our method is available at this https URL.

76.SSCAN: A Spatial-spectral Cross Attention Network for Hyperspectral Image Denoising ⬇️

Hyperspectral images (HSIs) have been widely used in a variety of applications thanks to the rich spectral information they are able to provide. Among all HSI processing tasks, HSI denoising is a crucial step. Recently, deep learning-based image denoising methods have made great progress and achieved great performance. However, existing methods tend to ignore the correlations between adjacent spectral bands, leading to problems such as spectral distortion and blurred edges in denoised results. In this study, we propose a novel HSI denoising network, termed SSCAN, that combines group convolutions and attention modules. Specifically, we use a group convolution with a spatial attention module to facilitate feature extraction by directing models' attention to band-wise important features. We propose a spectral-spatial attention block (SSAB) to exploit the spatial and spectral information in hyperspectral images in an effective manner. In addition, we adopt residual learning operations with skip connections to ensure training stability. The experimental results indicate that the proposed SSCAN outperforms several state-of-the-art HSI denoising algorithms.

77.Deep Learning Traversability Estimator for Mobile Robots in Unstructured Environments ⬇️

Terrain traversability analysis plays a major role in ensuring safe robotic navigation in unstructured environments. However, real-time constraints frequently limit the accuracy of online tests, especially in scenarios where realistic robot-terrain interactions are complex to model. In this context, we propose a deep learning framework, trained in an end-to-end fashion from elevation maps and trajectories, to estimate the occurrence of failure events. The network is first trained and tested in simulation over synthetic maps generated by the OpenSimplex algorithm. The prediction performance of the Deep Learning framework is illustrated by being able to retain over 94% recall of the original simulator at 30% of the computational time. Finally, the network is transferred and tested on real elevation maps collected by the SEEKER consortium during the Martian rover test trial in the Atacama desert in Chile. We show that transferring and fine-tuning of an application-independent pre-trained model retains better performance than training uniquely on scarcely available real data.

78.Orthogonal Ensemble Networks for Biomedical Image Segmentation ⬇️

Despite the astonishing performance of deep-learning based approaches for visual tasks such as semantic segmentation, they are known to produce miscalibrated predictions, which could be harmful for critical decision-making processes. Ensemble learning has shown to not only boost the performance of individual models but also reduce their miscalibration by averaging independent predictions. In this scenario, model diversity has become a key factor, which facilitates individual models converging to different functional solutions. In this work, we introduce Orthogonal Ensemble Networks (OEN), a novel framework to explicitly enforce model diversity by means of orthogonal constraints. The proposed method is based on the hypothesis that inducing orthogonality among the constituents of the ensemble will increase the overall model diversity. We resort to a new pairwise orthogonality constraint which can be used to regularize a sequential ensemble training process, resulting on improved predictive performance and better calibrated model outputs. We benchmark the proposed framework in two challenging brain lesion segmentation tasks --brain tumor and white matter hyper-intensity segmentation in MR images. The experimental results show that our approach produces more robust and well-calibrated ensemble models and can deal with challenging tasks in the context of biomedical image segmentation.

79.MIASSR: An Approach for Medical Image Arbitrary Scale Super-Resolution ⬇️

Single image super-resolution (SISR) aims to obtain a high-resolution output from one low-resolution image. Currently, deep learning-based SISR approaches have been widely discussed in medical image processing, because of their potential to achieve high-quality, high spatial resolution images without the cost of additional scans. However, most existing methods are designed for scale-specific SR tasks and are unable to generalise over magnification scales. In this paper, we propose an approach for medical image arbitrary-scale super-resolution (MIASSR), in which we couple meta-learning with generative adversarial networks (GANs) to super-resolve medical images at any scale of magnification in (1, 4]. Compared to state-of-the-art SISR algorithms on single-modal magnetic resonance (MR) brain images (OASIS-brains) and multi-modal MR brain images (BraTS), MIASSR achieves comparable fidelity performance and the best perceptual quality with the smallest model size. We also employ transfer learning to enable MIASSR to tackle SR tasks of new medical modalities, such as cardiac MR images (ACDC) and chest computed tomography images (COVID-CT). The source code of our work is also public. Thus, MIASSR has the potential to become a new foundational pre-/post-processing step in clinical image analysis tasks such as reconstruction, image quality enhancement, and segmentation.

80.Denoising Noisy Neural Networks: A Bayesian Approach with Compensation ⬇️

Noisy neural networks (NoisyNNs) refer to the inference and training of NNs in the presence of noise. Noise is inherent in most communication and storage systems; hence, NoisyNNs emerge in many new applications, including federated edge learning, where wireless devices collaboratively train a NN over a noisy wireless channel, or when NNs are implemented/stored in an analog storage medium. This paper studies a fundamental problem of NoisyNNs: how to estimate the uncontaminated NN weights from their noisy observations or manifestations. Whereas all prior works relied on the maximum likelihood (ML) estimation to maximize the likelihood function of the estimated NN weights, this paper demonstrates that the ML estimator is in general suboptimal. To overcome the suboptimality of the conventional ML estimator, we put forth an $\text{MMSE}{pb}$ estimator to minimize a compensated mean squared error (MSE) with a population compensator and a bias compensator. Our approach works well for NoisyNNs arising in both 1) noisy inference, where noise is introduced only in the inference phase on the already-trained NN weights; and 2) noisy training, where noise is introduced over the course of training. Extensive experiments on the CIFAR-10 and SST-2 datasets with different NN architectures verify the significant performance gains of the $\text{MMSE}{pb}$ estimator over the ML estimator when used to denoise the NoisyNN. For noisy inference, the average gains are up to $156%$ for a noisy ResNet34 model and $14.7%$ for a noisy BERT model; for noisy training, the average gains are up to $18.1$ dB for a noisy ResNet18 model.

81.Post-Radiotherapy PET Image Outcome Prediction by Deep Learning under Biological Model Guidance: A Feasibility Study of Oropharyngeal Cancer Application ⬇️

This paper develops a method of biologically guided deep learning for post-radiation FDG-PET image outcome prediction based on pre-radiation images and radiotherapy dose information. Based on the classic reaction-diffusion mechanism, a novel biological model was proposed using a partial differential equation that incorporates spatial radiation dose distribution as a patient-specific treatment information variable. A 7-layer encoder-decoder-based convolutional neural network (CNN) was designed and trained to learn the proposed biological model. As such, the model could generate post-radiation FDG-PET image outcome predictions with possible time-series transition from pre-radiotherapy image states to post-radiotherapy states. The proposed method was developed using 64 oropharyngeal patients with paired FDG-PET studies before and after 20Gy delivery (2Gy/daily fraction) by IMRT. In a two-branch deep learning execution, the proposed CNN learns specific terms in the biological model from paired FDG-PET images and spatial dose distribution as in one branch, and the biological model generates post-20Gy FDG-PET image prediction in the other branch. The proposed method successfully generated post-20Gy FDG-PET image outcome prediction with breakdown illustrations of biological model components. Time-series FDG-PET image predictions were generated to demonstrate the feasibility of disease response rendering. The developed biologically guided deep learning method achieved post-20Gy FDG-PET image outcome predictions in good agreement with ground-truth results. With break-down biological modeling components, the outcome image predictions could be used in adaptive radiotherapy decision-making to optimize personalized plans for the best outcome in the future.

82.Automatic calibration of time of flight based non-line-of-sight reconstruction ⬇️

Time of flight based Non-line-of-sight (NLOS) imaging approaches require precise calibration of illumination and detector positions on the visible scene to produce reasonable results. If this calibration error is sufficiently high, reconstruction can fail entirely without any indication to the user. In this work, we highlight the necessity of building autocalibration into NLOS reconstruction in order to handle mis-calibration. We propose a forward model of NLOS measurements that is differentiable with respect to both, the hidden scene albedo, and virtual illumination and detector positions. With only a mean squared error loss and no regularization, our model enables joint reconstruction and recovery of calibration parameters by minimizing the measurement residual using gradient descent. We demonstrate our method is able to produce robust reconstructions using simulated and real data where the calibration error applied causes other state of the art algorithms to fail.

83.High Throughput Soybean Pod-Counting with In-Field Robotic Data Collection and Machine-Vision Based Data Analysis ⬇️

We report promising results for high-throughput on-field soybean pod count with small mobile robots and machine-vision algorithms. Our results show that the machine-vision based soybean pod counts are strongly correlated with soybean yield. While pod counts has a strong correlation with soybean yield, pod counting is extremely labor intensive, and has been difficult to automate. Our results establish that an autonomous robot equipped with vision sensors can autonomously collect soybean data at maturity. Machine-vision algorithms can be used to estimate pod-counts across a large diversity panel planted across experimental units (EUs, or plots) in a high-throughput, automated manner. We report a correlation of 0.67 between our automated pod counts and soybean yield. The data was collected in an experiment consisting of 1463 single-row plots maintained by the University of Illinois soybean breeding program during the 2020 growing season. We also report a correlation of 0.88 between automated pod counts and manual pod counts over a smaller data set of 16 plots.

84.Hyper-Convolution Networks for Biomedical Image Segmentation ⬇️

The convolution operation is a central building block of neural network architectures widely used in computer vision. The size of the convolution kernels determines both the expressiveness of convolutional neural networks (CNN), as well as the number of learnable parameters. Increasing the network capacity to capture rich pixel relationships requires increasing the number of learnable parameters, often leading to overfitting and/or lack of robustness. In this paper, we propose a powerful novel building block, the hyper-convolution, which implicitly represents the convolution kernel as a function of kernel coordinates. Hyper-convolutions enable decoupling the kernel size, and hence its receptive field, from the number of learnable parameters. In our experiments, focused on challenging biomedical image segmentation tasks, we demonstrate that replacing regular convolutions with hyper-convolutions leads to more efficient architectures that achieve improved accuracy. Our analysis also shows that learned hyper-convolutions are naturally regularized, which can offer better generalization performance. We believe that hyper-convolutions can be a powerful building block in future neural network architectures solving computer vision tasks.

85.Prostate Gland Segmentation in Histology Images via Residual and Multi-Resolution U-Net ⬇️

Prostate cancer is one of the most prevalent cancers worldwide. One of the key factors in reducing its mortality is based on early detection. The computer-aided diagnosis systems for this task are based on the glandular structural analysis in histology images. Hence, accurate gland detection and segmentation is crucial for a successful prediction. The methodological basis of this work is a prostate gland segmentation based on U-Net convolutional neural network architectures modified with residual and multi-resolution blocks, trained using data augmentation techniques. The residual configuration outperforms in the test subset the previous state-of-the-art approaches in an image-level comparison, reaching an average Dice Index of 0.77.