Awesome

Video Person Re-ID: Fantastic Techniques and Where to Find Them

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}
}

[Update] Further worked (not published): Fine-Grained Re-Identification github || paper

Extension of Work done for Videos

CORRECTION

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

Specifications

Pytorch == 1.3.0
Python3

MODEL

RESULTS

LOSS

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

Performance

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

Model	mAP (re-rank)	CMC-1 (re-rank)	CMC-5 (re-rank)	CMC-20 (re-rank)
SOTA (w/o re-rank) (Fu et al.)[1]	81.2 (-)	86.2 (-)	95.7 (-)	- (-)
SOTA (with re-rank) (Fu et al.)[1]	80.8 (87.7)	86.3(87.2)	95.7(96.2)	98.1(98.6)
Baseline	76.7 (84.5)	83.3 (85.0)	93.8 (94.7)	97.4 (97.7)
Baseline + BOT	81.3 (88.4)	87.1 (87.6)	95.9 (96.0)	98.2 (98.4)
Baseline + BOT + OSM Loss	82.4 (88.1)	87.9 (87.6)	96.0 (95.7)	98.0 (98.5)
(Proposed) Baseline + BOT + OSM Loss + CL Centers	81.2 (88.5)	86.3 (88.0)	95.6 (96.1)	98.2 (98.5)
(Proposed) B-BOT + Attn-CL loss	82.9(87.8)	88.6(88.0)	96.2(95.4)	98.0(98.3)

PRID DATASET

Model	CMC-1	CMC-5	CMC-20
SOTA (Zeng, Tian, and Wu)[2]	96.1%	99.5	-
Baseline + BOT + OSM Loss + CL Centers	93.1	98.9	100.0
Baseline + BOT + OSM Loss + CL Centers (pretrained on MARS)	96.6	100	100

DATASET

MARS dataset:

MARS Dataset configuration (59th)

OSM LOSS hyper paramters

sigma = 0.9047814732165316
alpha = 2.8436551583293728
l = 0.5873389293193368

Triplet loss

margin = 4.4132437486402204e-05

(1 - beta_ratio) x Triplet_Loss + (beta_ratio) x OSM_Loss

beta_ratio = 1.0 (only osm loss used)

Optimizaer hyperparameters

gamma = 0.3282654557691594
weight_decay = 0.0005

NOTE

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

bag of tricks

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

hyper parameter optimization

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