light-reid

a toolbox of light reid for fast feature extraction and search

• light-model: model distillation (3x faster feature extraction)
• light-feature: binary code learning (6x faster retrieval)
• light-search: coarse2fine search (2x faster retrieval)

it features

• easy switch between light and non-light reid
• simple modules for reid implementation
• implementations of state-of-the-art deep reid models

What's New

• [2021.06]: we support vision transformers as cnn backbones. please refer base_config_duke_vit.yaml
• [2020.12]: we release a strong pipeline for occluded/partial re-id. please refer occluded_reid
• [2020.11]: we support pca_reduction to 128d with almost no accuracy drop. please refer bagtricks_pca
• [2020.11]: we support build with config files, making coding more simple. please refer bagtricks_buildwithconfigs
• [2020.08]: We release a toolbox of light-reid learning for faster inference, getting >30x faster speed.
• [2020.03]: We implement BagTricks and support IBN-Net, MSMT17, combineall, multi-dataset train. Please see branch version_py3.7_bot.
• [2019.03]: We give a clean implemention of BagTricks with python2.7. Please see branch version_py2.7.

Find our Works

• [2020.07]: [ECCV'20] Our work about Fast ReID has been accepted by ECCV'20. (Paper, Code)
• [2020.03]: [CVPR'20] Our work about Occluded ReID has been accepted by CVPR'20. (Paper, Code).
• [2020.01]: [AAAI'20] Our work about RGB-Infrared(IR) ReID has been accepted by AAAI'20. (Paper, Code).
• [2019.10]: [ICCV'19] Our work about RGB-Infrared(IR) ReID has been accepted by ICCV'19. (Paper, Code).
• [2019.05]: We implement PCB and achieve better performance than the offical one. (Code)

Installation

shellCopy
# clone this repo
git clone https://github.com/wangguanan/light-reid.git

# create environment
cd light-reid
conda create -n lightreid python=3.7
conda activate lightreid

# install dependencies
pip install -r requirements

# install torch and torchvision (select the proper cuda version to suit your machine)
conda install pytorch==1.4.0 torchvision -c pytorch
# install faiss for stable search
conda install faiss-cpu -c pytorch

Prepare Datasets

• download datasets that you need, reid_datasets.md lists various of datasets and their links.
• update datasets path at ./lightreid/data/datasets/datasetpaths.yaml

Quick Start

1 step to build a SOTA reid model with configs

Implemented reid methods and experimental results

• bagtricks_buildwithconfigs: easily implement a strong reid baseline
• bagtricks_pca: reduce feature dimension with PCA
• occluded_reid: a simple&strong reid baseline for occluded reid
• generalizable_reid: a reid model performs well on multiple datasets

Acknowledge

Our light-reid partially refers open-sourced
torch-reid and
fast-reid,
we thank their awesome contribution to reid community.

If you have any question about this reid toolbox, please feel free to contact me.
E-mail: guan.wang0706@gmail.com

LICENSE

light-reid is released released under the MIT License.

Citation

if you find the repo is useful, please kindly cite our works

@article{wang2020faster,
 title="Faster Person Re-Identification.",
 author="Guan'an {Wang} and Shaogang {Gong} and Jian {Cheng} and Zengguang {Hou}",
 journal="In Proceedings of the European Conference on Computer Vision (ECCV)",
 year="2020"
}

@article{wang2020honet,
 title="High-Order Information Matters: Learning Relation and Topology for Occluded Person Re-Identification.",
 author="Guan'an {Wang} and Shuo {Yang} and Huanyu {Liu} and Zhicheng {Wang} and Yang {Yang} and Shuliang {Wang} and Gang {Yu} and Erjin {Zhou} and Jian {Sun}",
 journal="In Proceedings of the 2020 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)",
 year="2020"
}