Pytorch implementation for M^3L

M3L

Learning to Generalize Unseen Domains via Memory-based Multi-Source Meta-Learning for Person Re-Identification (CVPR 2021)

Requirements

CUDA>=10.0
At least three 2080-Ti GPUs
Other necessary packages listed in requirements.txt
Training Data

The model is trained and evaluated on Market-1501, DukeMTMC-reID, MSMT17_V1, MSMT17_V2, CUHK03 and CUHK-NP

Unzip all datasets and ensure the file structure is as follow:
```
data    
│
└─── market1501 / dukemtmc / cuhknp / cuhk03 / msmt17v1 / msmt17v2
     │   
     └─── DukeMTMC-reID / Market-1501-v15.09.15 / detected / cuhk03_release / MSMT17_V1 / MSMT17_V2
```
Note:

In default, for CUHK03, we use the old protocol (CUHK03, 26,263 images of 1,367 IDs for training) as the source domain for training the model and the detected subset of the new protocol (CUHK-NP) as the target domain for testing; for MSMT17, we use the MSMT17_V2 for both training and testing.
We also provide the results of using the detected subset of CUHK-NP (7,365 images of 767 IDs for training) and MSMT17_V1 for both training and testing, and we recommend using this setting in future studies.

Run

ARCH=resMeta/IBNMeta
SRC1/SRC2/SRC3=market1501/dukemtmc/cuhk03/cuhknp/msmt17v1/msmt17v2
TARGET=market1501/dukemtmc/cuhknp/msmt17v1/msmt17v2

# train
CUDA_VISIBLE_DEVICES=0,1,2 python main.py \
-a $ARCH --BNNeck \
--dataset_src1 $SRC1 --dataset_src2 $SRC2 --dataset_src3 $SRC3 -d $TARGET \
--logs-dir $LOG_DIR --data-dir $DATA_DIR

# evaluate
python main.py \
-a $ARCH -d $TARGET \
--logs-dir $LOG_DIR --data-dir $DATA_DIR \
--evaluate --resume $RESUME

Results

m3l_results

You can download the above models in the paper from Google Drive. The model is named as $TARGET_$ARCH.pth.tar.

Acknowledgments

This repo borrows partially from MWNet,
ECN and
SpCL.

Citation

@inproceedings{zhao2021learning,
  title={Learning to Generalize Unseen Domains via Memory-based Multi-Source Meta-Learning for Person Re-Identification},
  author={Zhao, Yuyang and Zhong, Zhun and Yang, Fengxiang and Luo, Zhiming and Lin, Yaojin and Li, Shaozi and Nicu, Sebe},
  booktitle={CVPR},
  year={2021},
}