/Interleaved-Learning

Code for Style Interleaved Learning for Generalizable Person Re-identification （TMM 2023）

Primary LanguagePython

Style Interleaved Learning for Generalizable Person Re-identification (TMM 2023)

Requirements

  • CUDA>=10.0

  • At least four 1080-Ti GPUs

  • Training Data

    The model is trained and evaluated on Market-1501, DukeMTMC-reID, MSMT17_V1, CUHK03, CUHK-SYSU.

    Unzip all datasets and ensure the file structure is as follow:

    data    
│
└─── market1501 / dukemtmc / cuhk03 / msmt17v1 / cuhk_sysu
     │   
     └───  Market-1501-v15.09.15 / DukeMTMC-reID / cuhk03_release / MSMT17_V1 / CUHK-SYSU

Run

ARCH=resnet50
SRC1/SRC2/SRC3=market1501/dukemtmc/cuhk03/msmt17v1/cuhk_sysu
TARGET=market1501/dukemtmc/cuhk03/msmt17v1

# train baseline
CUDA_VISIBLE_DEVICES=0,1,2,3 python main.py \
-a resnet50 -b 64 --test-batch-size 256 --iters 200 --lr 3.5e-4 --epoch 70 \
--dataset_src1 msmt17v1 --dataset_src2 cuhk03 --dataset_src3 market1501 -d dukemtmc \
--logs-dir logs/Baseline \
--data-dir DATA_PATH

# train IL
CUDA_VISIBLE_DEVICES=0,1,2,3 python main.py \
-a resnet50 -b 64 --test-batch-size 256 --iters 200 --lr 3.5e-4 --epoch 70 \
--dataset_src1 msmt17v1 --dataset_src2 cuhk03 --dataset_src3 market1501 -d dukemtmc \
--logs-dir logs/IL \
--updateStyle \
--data-dir DATA_PATH

Note:

(1) The baseline setting in this code can be denoted as 'FBF' baseline, which is different with the reported 'FB' baseline in the paper. But they are similar in performance:

------------------------------------------------------------
|Baseline|→M       |→D       |→MS      |→C3      |Avg      |
|FB      |59.3/81.2|54.3/70.9|14.7/35.2|36.1/37.4|41.1/56.2|
|FBF     |59.7/81.5|53.6/73.0|13.7/33.4|35.3/35.9|40.6/56.0|
------------------------------------------------------------

(2) Just simply set '--updateStyle' can activate the interleaved learning.

Results

Acknowledgments

This repo borrows partially from M3L.

Citation

@article{tan2023style,
  title={Style Interleaved Learning for Generalizable Person Re-identification},
  author={Tan, Wentao and Ding, Changxing and Wang, Pengfei and Gong, Mingming and Jia, Kui},
  journal={IEEE Transactions on Multimedia},
  year={2023},
  publisher={IEEE},
  doi={10.1109/TMM.2023.3283878}
}

Contact

Email: ftwentaotan@mail.scut.edu.cn or 731584671@qq.com

如果可以当然还是希望用中文contact我啦！