ReID-Survey with a Powerful AGW Baseline
Deep Learning for Person Re-identification: A Survey and Outlook. arXiv
An implementation of AGW for cross-modality visible-infrared Re-ID is HERE.
Highlights
-
A comprehensive survey with in-depth analysis for person Re-ID in recent years (2016-2019).
-
A new evaluation metric, namely mean Inverse Negative Penalty (mINP), which measures the ability to find the hardest correct match.
-
A new AGW baseline with non-local Attention block, Generalized mean pooling and Weighted regularization triplet. It acheieves competitive performance on both single-modality and cross-modality Re-ID tasks.
AGW on Single-Modality Re-ID with mINP
DukeMTMC dataset
| Method | Pretrained | Rank@1 | mAP | mINP | Model | Paper |
|---|---|---|---|---|---|---|
| BagTricks | ImageNet | 86.4% | 76.4% | 40.7% | Code | Bag of Tricks and A Strong Baseline for Deep Person Re-identification. In ArXiv 19. PDF |
| ABD-Net | ImageNet | 89.0% | 78.6% | 42.1% | Code | ABD-Net: Attentive but Diverse Person Re-Identification. In ICCV 19. PDF |
| AGW | ImageNet | 89.0% | 79.6% | 45.7% | GoogleDrive | Deep Learning for Person Re-identification: A Survey and Outlook |
Market-1501 dataset
| Method | Pretrained | Rank@1 | mAP | mINP | Model | Paper |
|---|---|---|---|---|---|---|
| BagTricks | ImageNet | 94.5% | 85.9% | 59.4% | Code | Bag of Tricks and A Strong Baseline for Deep Person Re-identification. In ArXiv 19. arXiv |
| ABD-Net | ImageNet | 95.6% | 88.3% | 66.2% | Code | ABD-Net: Attentive but Diverse Person Re-Identification. In ICCV 19. PDF |
| AGW | ImageNet | 95.1% | 87.8% | 65.0% | GoogleDrive | Deep Learning for Person Re-identification: A Survey and Outlook. In ArXiv 20. arXiv |
CUHK03 dataset
| Method | Pretrained | Rank@1 | mAP | mINP | Model | Paper |
|---|---|---|---|---|---|---|
| BagTricks | ImageNet | 58.0% | 56.6% | 43.8% | Code | Bag of Tricks and A Strong Baseline for Deep Person Re-identification. In ArXiv 19. PDF |
| AGW | ImageNet | 63.6% | 62.0% | 50.3% | GoogleDrive | Deep Learning for Person Re-identification: A Survey and Outlook. In ArXiv 20. arXiv |
MSMT17 dataset
| Method | Pretrained | Rank@1 | mAP | mINP | Model | Paper |
|---|---|---|---|---|---|---|
| BagTricks | ImageNet | 63.4% | 45.1% | 12.4% | Code | Bag of Tricks and A Strong Baseline for Deep Person Re-identification. In ArXiv 19. arXiv |
| AGW | ImageNet | 68.3% | 49.3% | 14.7% | GoogleDrive | Deep Learning for Person Re-identification: A Survey and Outlook. In ArXiv 20. arXiv |
Quick Start
1. Prepare dataset
Create a directory to store reid datasets under this repo, taking Market1501 for example
cd ReID-Survey
mkdir toDataset
-
Set
_C.DATASETS.ROOT_DIR = ('./toDataset')inconfig/defaults.py -
Download dataset to toDataset/ from http://www.liangzheng.org/Project/project_reid.html
-
Extract dataset and rename to
market1501. The data structure would like:
toDataset
market1501
bounding_box_test/
bounding_box_train/
......
2. Train
To train a AGW model with on Market1501 with GPU device 0, run similarly:
python3 tools/main.py --config_file='configs/AGW_baseline.yml' MODEL.DEVICE_ID "('0')" DATASETS.NAMES "('market1501')" OUTPUT_DIR "('./log/market1501/Experiment-AGW-baseline')"
3. Test
To test a AGW model with on Market1501 with weight file './pretrained/dukemtmc_AGW.pth', run similarly:
python3 tools/main.py --config_file='configs/AGW_baseline.yml' MODEL.DEVICE_ID "('0')" DATASETS.NAMES "('market1501')" MODEL.PRETRAIN_CHOICE "('self')" TEST.WEIGHT "('./pretrained/market1501_AGW.pth')" TEST.EVALUATE_ONLY "('on')" OUTPUT_DIR "('./log/Test')"
Citation
Please kindly cite this paper in your publications if it helps your research:
@article{arxiv20reidsurvey,
title={Deep Learning for Person Re-identification: A Survey and Outlook},
author={Ye, Mang and Shen, Jianbing and Lin, Gaojie and Xiang, Tao and Shao, Ling and Hoi, Steven C. H.},
journal={arXiv preprint arXiv:2001.04193},
year={2020},
}
Contact: [email protected]
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent