A baseline for clothes-change Person ReID task.
This code is implemented based on JD Fast-reid.
- [2020.12.15] LTCC dataset is supported.
- [2020.12.16] Add evaluation code for clothes-change setting.
- [2020.12.16] Train without/with clothes labels.
- Add human keypoints
- Add human mask
See INSTALL.md.
The clothes-change datasets include:
- LTCC [dataset] [paper] : 17119 images from 152 identities
- VC-Clothes [dataset] [paper]: 19060 images from 512 identities (Virtual) + 4324 images from 28 identities (Real)
- PRCC [dataset] [paper]: 33698 images from 221 identities
Our experiments are based on LTCC. You can place LTCC dataset to make the data folder like:
${ROOT}
|-- datasets
`-- |-- LTCC_ReID
`-- |-- train
|--- test
|--- query
`--- info
|-- Other_Dataset
Train:
Single-GPU
python tools/train_net.py --config-file ./configs/LTCC/bagtricks_R50.yml MODEL.DEVICE "cuda:0"
Multi-GPU
python tools/train_net.py --config-file ./configs/LTCC/bagtricks_R50.yml --num-gpus 4
Test:
Standard Setting: The images with the same identity and the same camera view are discarded during testing.
python tools/train_net.py --config-file ./configs/LTCC/bagtricks_R50.yml --eval-only \
MODEL.WEIGHTS /path/to/checkpoint_file MODEL.DEVICE "cuda:0"
Cloth-changing Setting: The images with same identity, camera view and clothes are discarded during testing.
python tools/train_net.py --config-file ./configs/LTCC/bagtricks_R50.yml --eval-only --cconly \
MODEL.WEIGHTS /path/to/checkpoint_file MODEL.DEVICE "cuda:0"
For more options, see python ./tools/train_net.py -h
.
We provide some baseline results and trained models available for download:
These models are trained for 120 epochs with batch size=64 on 2 GeForce GTX TITAN X.
Method | Backbone | Standard | Cloth-changing | download | ||||
---|---|---|---|---|---|---|---|---|
Rank@1 | Rank@5 | mAP | Rank@1 | Rank@5 | mAP | |||
Baseline | Res-50 | 67.55% | 77.48% | 32.64% | 33.93% | 49.49% | 15.57% | - |
Baseline(w/ clo) | Res-50 | 73.43% | 81.74% | 38.54% | 31.89% | 48.47% | 15.47% | Model |
The code is released under the Apache 2.0 license.