Official Pytorch Implementation of the paper: Paper
Video Person Re-ID: Fantastic Techniques and Where to Find Them (Accepted in AAAI'20)
Priyank Pathak, Amir Erfan Eshratifar, Michael Gormish
(Work done in collaboration of NYU and Clarifai)
If you find this useful, feel free to cite this

@misc{pathak2019video,
    title={Video Person Re-ID: Fantastic Techniques and Where to Find Them},
    author={Priyank Pathak and Amir Erfan Eshratifar and Michael Gormish},
    year={2019},
    eprint={1912.05295},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

Extension of Work done for Videos

• Revisiting Temporal Modeling for Video-based Person ReID || paper
• Bag of Tricks and A Strong ReID Baseline || paper
• Online-Soft-Mining-and-Class-Aware-Attention || paper

PRID results have been reported for split 0. Please consider the current SOTA as R-1 : 95.5 R-5: 100 R10: 100 for Global-local temporal representations for video person re-identification. In ICCV, 2019

Pytorch == 1.3.0
Python3

loss = ide_loss + (1 - beta_ratio)*triplet_loss + center_loss * cetner_loss_weight + beta_ratio * osm_caa_loss + attention_loss

Use 59 configuration in cl_centers.conf and vals.conf for MARS DATASET and 27 and 24 configuration in prid.conf for prid DATASET

MARS DATASET

PRID DATASET

MARS dataset:

• sigma = 0.9047814732165316
• alpha = 2.8436551583293728
• l = 0.5873389293193368

• margin = 4.4132437486402204e-05

• beta_ratio = 1.0 (only osm loss used)

• gamma = 0.3282654557691594
• weight_decay = 0.0005

Change the dataset path in tools/data_manager.py line 21 : storage_dir = "/beegfs/pp1953/"
(apologies for the inconvenience)

normal architecrture : args.arch = "ResNet50ta_bt"
For attention loss : args.arch = "ResNet50ta_bt2"

python bagoftricks.py --name="_CL_CENTERS_" --validation-training --cl-centers --opt=8
python bagoftricks.py --name="_triplet_OSM_only_" --validation-training --use-OSMCAA --opt=8
python bagoftricks.py --name="_triplet_only_" --validation-training --opt=8
python bagoftricks.py --name="_ilidsvid_" --validation-training --opt=8
python bagoftricks.py --name="_prid_" --validation-training --opt=24
python config_trainer.py --focus=map --dataset=mars --opt=59 --name=_mars_attncl_centers_ --cl-centers --attn-loss

python hyper_supervise_validation.py --focus="map" --opt=8
python hyper_supervise_validation.py --focus="rerank_map" --opt=8
python hyper_supervise_validation.py --focus="map" --sampling="inteliigi" --opt=8
python hyper_supervise_validation.py --focus="rerank_map" --sampling="inteliigi" --opt=8

Ref:

[1] STA: Spatial-Temporal Attention for Large-Scale Video-based Person Re-Identification
[2] Person Re-identification with Hierarchical Deep Learning Feature and efficient XQDA Metric