[ECCV 2022] Mimic Embedding via Adaptive Aggregation: Learning Generalizable Person Re-identification [paper]
2022-07-14: Update Code.
2022-08-03: Update Paper.
- Requirements: Market1501, CUHK03, CUHK-SYSU, MSMT17_v2
Please put all the datasets in one directory, and modify DATASETS in projects/META/configs/Base-cnn.yml
- Please check fast-reid for fast-reid installation
- Please check Apex for Apex installation
- Compile with cython to accelerate evalution:
bash cd fastreid/evaluation/rank_cylib; make all
If you want to train with 1-GPU, run:
python projects/META/train_net.py --config-file projects/META/configs/r50.yml MODEL.DEVICE "cuda:0"
If you want to train with 4-GPU, run:
CUDA_VISIBLE_DEVICES=0,1,2,3 python projects/META/train_net.py --config-file projects/META/configs/r50.yml --num-gpus 4
You can get the results in our paper by training with 4-GPU, please modify SOLVER.IMS_PER_BATCH in projects/META/configs/Base-cnn.yml (64 for 1-GPU and 256 for 4-GPU)
To evaluate a model's performance, use:
python projects/META/train_net.py --config-file projects/META/configs/r50.yml --eval-only MODEL.WEIGHTS /path/to/checkpoint_file MODEL.DEVICE "cuda:0"
If you have any question about the project, please feel free to contact me.
E-mail: [email protected]
The code was developed based on the ’fast-reid’ toolbox https://github.com/JDAI-CV/fast-reid.