This project provides a simple but strong Re-ID pipeline for occluded/partial Re-ID.
The pipeline achieves very high accuracy on three popular occluded/partial datasets, includes Occluded-ReID, Partial-iLIDS, Partial-ReID.
Our CVPR2020 work HONet is based on this pipeline and achieves better accuracy, please refer its github for more details.
Different from common Re-ID which assume query and gallery images are holistic (e.g. head, body, legs are visible), occluded/partial Re-ID is more general which accepts partial/occluded images (only partial region is visible and the others are invisible due to outlier our occlusion) as queries.
-
Please download Market-1501, Occluded-ReID, Partial-ReID, Partial-iLIDs. Links can be found here.
-
Please downloaed the trained pose model pose_hrnet_w48_256x192.pth and set yaml files
model.head.pose_model_pathto be your own path.
# train
python train.py --config_file ./config_occludedreid.yaml
# infer
python infer.py --config_file ./config_occludedreid.yaml --model_path /path/to/model.pth
Settings (on a MacBook Pro (Retina, 13-inch, Mid 2014))
- GPU: TITAN XP (memory 12194MB)
- CPU: 2.6 GHz Dual-Core Intel Core i5
- Memory: 8 GB 1600 MHz DDR3
| Methods | Backbone | Conf. | Occluded-ReID | Partial-ReID | Partial-iLIDs | Github/Model |
|---|---|---|---|---|---|---|
| OONet(Ours) | ResNet50 | - | 72.1(64.0) | 86.3(90.0) | 70.6(82.0) | model |
| OONet(Ours) | ResNet50-ibna | - | 78.7(70.9) | 85.0(90.1) | 73.9(83.0) | model |
| HONet | ResNet50 | CVPR2020 | 80.3(70.2) | 85.3(91.0) | 72.6(86.4) | github |
| TCSDO | ResNet50 | ArXiv2019 | 73.7(67.9) | 82.7(-) | - | - |
| FPR | ResNet50 | CVPR2019 | 78.3(68.0) | 81.0(-) | 68.1(-) | - |
| PGFA | ResNet50 | ICCV2019 | - | 68.0(80.0) | 69.1(80.9) | - |
| VPM | ResNet50 | ICCV2019 | - | 67.7(81.9) | 65.5(74.8) | - |
| DSR | ResNet50 | CVPR2018 | 72.8(62.8) | 50.7(70.0) | 58.8(67.2) | github |