The official implementation for the A High-Accuracy Unsupervised Person Re-identification Method Using Auxiliary Information Mined from Datasets.
git clone https://github.com/tenghehan/AuxUSLReID.git
cd AuxUSLReID
conda create -n your_env_name python=3.7.9
conda install –yes –file requirements.txt
python setup.py develop
Download the raw datasets DukeMTMC-VideoReID, MARS and then unzip them under the directory like
AuxUSLReID/datasets
├── DukeMTMC-VideoReID
└── MARS
# Train on MARS
python scripts/main.py --config-file configs/USL_MTCluster.yml
# Train on DukeMTMC-VideoReID
python scripts/main.py --config-file configs/USL_MTCluster_Duke.yml# Test on MARS
python scripts/main.py --config-file configs/USL_MTCluster.yml RESUME 'log/MTCluster_GPU4/model_mAP_best.pth.tar' TEST.EVAL_ONLY True
# Test on DukeMTMC-VideoReID
python scripts/main.py --config-file configs/USL_MTCluster_Duke.yml RESUME 'log/MTCluster_GPU4/model_mAP_best.pth.tar' TEST.EVAL_ONLY TrueThe parameters are set in configs/USL_MTCluster(_Duke).yml and unreidusl/config/defaults.py
Auxiliary information exploiting modules TOC, STS and SCP can only be used for DukeMTMC-VideoReID. The pickle files related to the three modules are placed in AuxUSLReID/pickles and are generated from DukeMTMC-VideoReID.
The parameters related to the methods we proposed are introducted as follows. Modify the config (config/*.yml) as our paper suggests before training.
# 1:CE 2:LSCE 3:RLSCE
_C.MODEL.LOSSES.ID_OPTION = 3
# K in RLSCE
_C.MODEL.LOSSES.ID_K = 60
# epsilon in RLSCE
_C.MODEL.LOSSES.CE.EPSILON = 0.1
# 1:TL 2:WATL
_C.MODEL.LOSSES.TRI_OPTION = 1
# margin in WATL
_C.MODEL.LOSSES.TRI_MARGIN = 0.5
# lambda in WATL
_C.MODEL.LOSSES.TRI_LAMBDA = -1.0# -1.: fixed training iteration
# otherwise: iterations = pid * ITERS_CLUSTER_MULTIPLIER
_C.DATALOADER.ITERS_CLUSTER_MULTIPLIER = 0.6# whether to use TOC in training
_C.CLUSTER.ICDBSCAN_OPTION = False
# TOC calculated for training samples
_C.CLUSTER.ICDBSCAN_CONSTRAINT_PATH = 'cannotlink_pts.pickle'
# whether to use TOC in inference
_C.TEST.OVERLAP_CONSTRAINT = False
# TOC calculated for query and gallery
_C.TEST.CONSTRAINT_PATH = 'overlap_relation.pickle'
# whether to use STS in training
_C.CLUSTER.STR_OPTION = False
# lambda and gamma in STS
_C.CLUSTER.STR_LAMBDA = 0.6
_C.CLUSTER.STR_GAMMA = 8.0
# STS calculated for training samples
_C.CLUSTER.STR_PATH = "spatio_temporal_relation_train.pickle"
# whether to use STS in inference
_C.TEST.STR_OPTION = False
# STS calculated for query and gallery
_C.TEST.STR_PATH = "spatio_temporal_relation_infer.pickle"
# whether to use SCP in training
_C.CLUSTER.CCE_OPTION = False
# lambda in SCP
_C.CLUSTER.CCE_LAMBDA = 0.006
# SCP calculated for training samples
_C.CLUSTER.CCE_PATH = "cross_camera_relation_train.pickle"
# whether to use SCP in inference
_C.TEST.CCE_OPTION = False
# SCP calculated for query and gallery
_C.TEST.CCE_PATH = "cross_camera_relation_infer.pickle"The model weights of Ours and Ours- can be download from the link
If you find this code useful for your research, please cite our paper
@misc{teng2022highaccuracy,
title={A High-Accuracy Unsupervised Person Re-identification Method Using Auxiliary Information Mined from Datasets},
author={Hehan Teng and Tao He and Yuchen Guo and Guiguang Ding},
year={2022},
eprint={2205.03124},
archivePrefix={arXiv},
primaryClass={cs.CV}
}