Part Convolutional Baseline

This project implements PCB (Part-based Convolutional Baseline) of paper Beyond Part Models: Person Retrieval with Refined Part Pooling using pytorch.

Dependency

• python 3.6
• pytorch 1.0
• torchvision
• ignite
• yacs

Usage

python3 train.py

python3 eval.py [gpu-id] [chekpoint-path] 

Structure

├── configs
│   ├── config.yml
│   └── default.py
├── data
│   ├── dataset.py
│   └── __init__.py
├── engine
│   ├── create_reid_engine.py
│   ├── __init__.py
│   └── scalar_metric.py
├── eval.py
├── extract.py
├── layers
│   ├── am_softmax.py
│   └── norm_linear.py
├── models
│   ├── __init__.py
│   └── pcb.py
├── README.md
├── solver
│   └── lr_scheduler.py
├── train.py
├── transform
│   └── random_erase.py
└── utils
    ├── evaluation.py
    └── initializer.py

Performance

DukeMTMC-reID

setting	mAP	Rank-1
paper-1536	65.30	81.90
share-embed-1536	70.15	84.38
independent-embed-1536	71.48	85.10
paper-12288	66.10	81.70
share-embed-12288	64.91	82.90
independent-embed-12288	65.26	83.98

Market-1501

setting	mAP	Rank-1
paper-1536	77.30	92.40
share-embed-1536	78.68	92.96
independent-embed-1536	79.79	93.08
paper-12288	77.40	92.30
share-embed-12288	73.09	92.36
independent-embed-12288	72.72	92.07

MSMT17

setting	mAP	Rank-1
share-embed-1536	42.39	69.59
independent-embed-1536	47.20	74.05

We also evaluate the original setting with a decay step-size of 20 which is adopted in some re-implementations.

DukeMTMC-reID

setting	mAP	Rank-1
paper-1536	65.30	81.90
share-embed-1536	67.95	82.76
independent-embed-1536	71.70	84.91
paper-12288	66.10	81.70
share-embed-12288	62.85	81.28
independent-embed-12288	65.51	83.39

Market-1501

setting	mAP	Rank-1
paper-1536	77.30	92.40
share-embed-1536	78.03	92.93
independent-embed-1536	80.30	93.26
paper-12288	77.40	92.30
share-embed-12288	72.29	91.74
independent-embed-12288	72.29	91.95

MSMT17

setting	mAP	Rank-1
share-embed-1536	42.45	70.35
independent-embed-1536	46.75	72.96

Ablation Studies

Experiments with shared 1x1 conv on DukeMTMC-reID

num parts	mAP-h	R1-h	mAP-g	R1-g
1	58.38	76.88	59.80	80.20
2	66.68	82.54	62.84	81.41
3	66.66	83.48	63.15	81.86
4	66.96	82.40	62.21	81.28
5	67.17	82.67	61.57	81.10
6	70.09	84.20	64.90	83.21

Experiments with unshared 1x1 conv on DukeMTMC-reID

num parts	mAP-h	R1-h	mAP-g	R1-g
1	58.38	76.88	59.80	80.20
2	66.34	83.03	61.84	81.32
3	67.80	82.89	60.60	80.38
4	70.07	84.42	63.40	82.40
5	69.74	84.38	63.92	82.63
6	71.48	85.10	65.26	83.98