Deep Constrained Dominant Sets for Person Re-Identification (DCDS)
Pytorch Implementation for our ICCV2019 work, Deep Constrained Dominant Sets for Person Re-Identification. This implementation is based on open-reid and kpm_rw_person_reid.
- python 2.7
- PyTorch (we run the code under version 0.3.0)
- metric-learn 0.3.0
- torchvision 0.2.1
- cd DCDS
- install setup.py
For single dataset (SD) setup we use Market1501, CUHK03, and DukeMTMC datasets.
cd examples/
mkdir data
cd data/
mkdir market1501
cd market1501
mkdir raw/
mv dir_of_market1501_zip raw/
Repeate this for CUHK03 and Dukemtmc.
python examples/main.py -d market1501 -b 64 -a resnet101 --features 2048 --lr 0.0001 --ss 10 --epochs 100 --dropout 0 --weight-decay 0 --logs-dir examples/logs/market1501-final-model
python examples/main.py -d dukemtmc -b 64 -a resnet101 --features 2048 --lr 0.0001 --ss 10 --epochs 100 --dropout 0 --weight-decay 0 --logs-dir examples/logs/dukemtmc-final-model
- Same for CUHK03
Download the trained models
python examples/main.py -d market1501 -b 64 -a resnet101 --features 2048 --evaluate --evaluate-from examples/logs/market1501-final-model/model_best.pth.tar
python examples/main.py -d dukemtmc -b 64 -a resnet101 --features 2048 --evaluate --evaluate-from examples/logs/dukemtmc-final-model/model_best.pth.pth.tar
Market-1501 | CUHK03 | DukeMTMC-reID | |||||||
---|---|---|---|---|---|---|---|---|---|
mAP(%) | rank-1 | rank-5 | mAP(%) | rank-1 | rank-5 | mAP(%) | rank-1 | rank-5 | |
DCDS (SD) | 81.5 | 92.9 | 97.4 | 90.7 | 93.3 | 99.1 | 70.3 | 83.6 | 90.4 |
@InProceedings{Alemu_2019_ICCV,
author = {Alemu, Leulseged Tesfaye and Pelillo, Marcello and Shah, Mubarak},
title = {Deep Constrained Dominant Sets for Person Re-Identification},
booktitle = {The IEEE International Conference on Computer Vision (ICCV)},
month = {October},
year = {2019}
}