PROCA: Place Recognition under Occlusion and Changing Appearance via Disentangled Representations

Paper | Video

Yue Chen¹, Xingyu Chen¹†⚑, Yicen Li²

†Corresponding Author, ⚑Project Lead, ¹Xi'an Jiaotong University, ²McMaster University

This repository is an official implementation of PROCA using pytorch.

Usage

Prerequisites

• Python 3
• Pytorch and torchvision (https://pytorch.org/)
• TensorboardX

Install

• Clone this repo:

git clone https://github.com/rover-xingyu/PROCA.git
cd PROCA

Datasets

• The dataset we use in this paper is from CMU-Seasons Dataset.
• In order to verify the generalization ability of PROCA, we sample images from the urban part as the training set, and evaluate on the suburban and park parts.
• We label the images with occlusion and without occlusion depending on if there are dynamic objects in the images. You can find our labels in dataset/CMU_Seasons_Occlusions.json
• The dataset is organized as follows:

    ├── CMU_urban
    │   ├── trainA // images with appearance A without occlusion
    │   │   ├── img_00119_c1_1303398474779487us_rect.jpg
    │   │   ├── ...
    │   ├── trainAO // images with appearance A with occlusion
    │   │   ├── img_00130_c0_1303398475779409us_rect.jpg
    │   │   ├── ...
    │   ├── ...
    │   │   trainL // images with appearance L without occlusion
    │   │   ├── img_00660_c0_1311874734447600us_rect.jpg
    │   │   ├── ...
    │   ├── trainLO // images with appearance L with occlusion
    │   │   ├── img_00617_c1_1311874730447615us_rect.jpg
    │   │   ├── ...

Citation

If you find this project useful for your research, please use the following BibTeX entry.

@inproceedings{chen2023place,
  title={Place Recognition under Occlusion and Changing Appearance via Disentangled Representations},
  author={Chen, Yue and Chen, Xingyu and Li, Yicen},
  booktitle={2023 IEEE International Conference on Robotics and Automation (ICRA)},
  pages={1882--1888},
  year={2023},
  organization={IEEE}
}

Acknowledge

Our code is based on the awesome pytorch implementation of Diverse Image-to-Image Translation via Disentangled Representations (DRIT++ and MDMM). We appreciate all the contributors.