UAD: Localization Uncertainty Estimation for Anchor-Free Object Detection

This repository inlcudes official implementations and model weights for UAD.
[Arxiv][BibTeX]

C_L, C_R, C_T, and C_B denote the estimated certainty in [0, 1] values with respect to left, right, top, and bottom. For example, the proposed UAD estimates low top-directional certainty due to its ambiguous head boundary of the cat wearing a hat. This demonstrates that our method enables the detection network to quantify which direction is uncertain due to unclear or ambiguous regions of objects.

Estimated uncertainty examples of the proposed UAD. Since there is no supervision of uncertainty, we analyze the estimated uncertainty qualitatively. UAD estimates lower certainties on unclear and ambiguous boundaries due to occlusion and shadow. For example, Both the surfboard and the person in the left-bottom image have much lower bottom-directional certainties (i.e., C_B : 55% and 22%) as their shapes are quite unclear due to the water. Also, the right-directional certainty (C_R) of the woman in the right-bottom image is estimated only by 1% because it is covered by the tail of a cat on the TV.

Models

Name	box AP (val)	box AP (test-dev)	download
UAD_R_50_ms_2x	42.2	42.4	model
UAD_R_101_ms_2x	44.0	44.2	model
UAD_X_101_64x4d_ms_2x	45.2	45.4	model
UAD_X_101_32x8d_dcnv2_ms_2x	48.0	48.4	model

Installation

First install Detectron2 following the official guide: INSTALL.md. Then build UAD with:

git clone https://github.com/youngwanLEE/UAD.git
cd UAD
python setup.py build develop

How to demo

1. Pick a model and its config file, for example, uad_r_50_ms_2x.yaml.
2. Download the pretrained model using the above download link.
3. Run the demo with

python demo/demo.py \
    --config-file configs/uad/uad_r_50_ms_2x.yaml \
    --input test_images/000000580743.jpg \
    --output ./ \
    --opts MODEL.WEIGHTS uadet_r_50_ms_2x_42_23.pth

--config-file: config file for model
--input: input images/videos path
--output: output directory to save the output
--opts MODEL.WEIGHT: the pretrained weight path

How to train on COCO

We train UAD models using a machine equipped with 8 GPUs.If you want to change the number of GPUs, use --num-gpus argument.
Detectron2's document may help you for more details.

python tools/train_net.py \
    --config-file configs/uad/uad_r_50_ms_2x.yaml \
    --num-gpus 8 \
    OUTPUT_DIR training_dir/uad_r_50_ms_2x

To evaluate the model after training, run:

python tools/train_net.py \
    --config-file configs/uad/uad_r_50_ms_2x.yaml \
    --eval-only \
    --num-gpus 8 \
    OUTPUT_DIR training_dir/uad_r_50_ms_2x \
    MODEL.WEIGHTS training_dir/uad_r_50_ms_2x/model_final.pth

License

This project is under CC-BY-NC 4.0 license. Please see LICENSE for details.

Acknowledgement

This implementation are based on Detectron2 library.
This work was supported by Institute of Information & Communications Technology Planning & Evaluation(IITP) grant funded by the Korea government(MSIT) (No.2014-3-00123, Development of High Performance Visual BigData Discovery Platform for Large-Scale Realtime Data Analysis, No.2022-0-00124, Development of Artificial Intelligence Technology for Self-Improving Competency-Aware Learning Capabilities).

Citing UAD

@inproceedings{lee2022uad,
      title={Localization Uncertainty Estimation for Anchor-Free Object Detection}, 
      author={Youngwan Lee and Jong-Won Hwang and Hyung-Il Kim and Kimin Yun and Yongjin Kown and Sung Ju Hwang},
      booktitle={ECCV-Workshop}
      year={2022},
}