The official repository for Hybrid Dynamic Contrast and Probability Distillation for Unsupervised Person Re-Id.
HDCPD
achieves state-of-the-art performances on both unsupervised learning tasks and unsupervised domain adaptation tasks for person re-ID.
For now, we could only release the test code and model already trained. We will gradually release our training code as the paper delivered.
We use 4 datasets in our train and test, including Market1501, DukeMTMC-ReID, MSMT17, PersonX. Please unzip the datasets under the diretory like
HDCPD-ReID/data
├── market1501
│ └── Market-1501-v15.09.15
├── msmt17
│ └── MSMT17_V1
├── personx
│ └── PersonX
└── duke
└── DukeMTMC-reID
We utilize 2 tesla v100 GPU for training
To recover UDA results, run:
sh uda.sh
To recover USL results, run:
sh usl.sh
We utilize 1 tesla v100 GPU for testing.
You can download model for market1501, duke and personx from Here with access code estr
To evaluate the model released, run:
CUDA_VISIBLE_DEVICES=0 \
python test.py -d $DATASET \
--resume $PATH_OF_MODEL
To recover "Ours+ClusterContrast" results, see "ours+cc" profiles.
Datasets | mAP(%) | R@1(%) | R@5(%) | R@10(%) |
---|---|---|---|---|
Market1501 | 81.6 | 92.6 | 97.4 | 98.2 |
DukeMTMC | 69.0 | 82.9 | 90.9 | 93 |
PersonX | 84.1 | 94.4 | 98.7 | 99.5 |
MSMT17 | 24.6 | 50.2 | 61.4 | 65.7 |
Thanks to SpCL. It is an excellent USL framework and deeply inspires our work.