Video-based Person Re-identification without Bells and Whistles

[Paper] [arXiv] [video]

Chih-Ting Liu, Jun-Cheng Chen, Chu-Song Chen and Shao-Yi Chien,
Analysis & Modeling of Faces & Gestures Workshop jointly with IEEE Conference on Computer Vision and Pattern Recognition (CVPRw), 2021

This is the pytorch implementatin of Coarse-to-Fine Axial Attention Network (CF-AAN) for video-based person Re-ID.
It achieves 91.3% in rank-1 accuracy and 86.5% in mAP on our aligned MARS dataset.

News

2021-06-13:

• We release the code and aligned dataset for our work.
• We update the Readme related to our new dataset, and the others will be updated gradually.

2021-06-18:

• We update the description for training and testing CF-AAN.

Aligned dataset with our re-Detect and Link module

Download Link :

• MARS (DL) : [Google Drive]
• For DukeV, we didn't perform DL on DukeMTMC-VideoReID because the bounding boxes are greound truth annotations.

Results

The video tracklet will be re-Detected, linked (tracking) and padded to the original image size, as follow.

Folder Structure

MARS dataset:

MARS-DL/
|-- bbox_train/
|-- bbox_test/
|-- info/
|-- |-- mask_info.csv (for DL mask)
|-- |-- mask_info_test.csv  (for DL mask)
|-- |-- clean_tracks_test_info.mat (for new evaluation protocol)
|-- |-- .... (other original info files)

DukeV dataset:

DukeMTMC-VideoReID/
|-- train/
|-- gallery/
|-- query/

You can put this two folders under your root dataset directory.

path to your root dir/
|-- MARS-DL/
|-- DukeMTMC-VideoReID/

Coarse-to-Fine Axial Attention Network (CF-AAN)

Requirement

We use Python 3.6, Pytorch 1.5 and Pytorch-ignite in this project. To install required modules, run:

pip3 install -r requirements.txt

Training

Train CF-AAN on MARS-DL

You can alter the argument in scripts/AA_M.sh and run it with:

sh scripts/AA_M.sh

Or, you can directly type:

python3 tools/train.py --config_file='configs/video_baseline.yml' MODEL.DEVICE_ID "('0,1')" DATASETS.NAMES "('mars',)" INPUT.SEQ_LEN 6 \
                                                                   OUTPUT_DIR "./ckpt_DL_M/MARS_DL_s6_resnet_axial_gap_rqkv_gran4" SOLVER.SOFT_MARGIN True \
                                                                   MODEL.NAME 'resnet50_axial' MODEL.TEMP 'Done' INPUT.IF_RE True \
                                                                   DATASETS.ROOT_DIR '<PATH TO DATASET ROOT DIRECTORY>'

* <PATH TO DATASET ROOT DIRECTORY> is the directory containing both MARS and DukeV dataset.

Train Non-local or baseline on MARS

You can alter the argument in scripts/NL_M.sh & scripts/baseline_M.sh and run it with:

sh scripts/AA_M.sh & sh scripts/baseline_M.sh

Train models on DukeMTMC-VideoReID

You can use the scripts scripts/AA_D.sh, scripts/NL_D.sh, & scripts/baseline_D.sh

Notes

If you want to train on original MARS dataset, you just need to change the comment in data/datasets/MARS.py :

class MARS(BaseVideoDataset):
    dataset_dir = 'MARS'
    # dataset_dir = 'MARS-DL'
    info_dir = 'info

Testing

You can alter the argument in scripts/test_M.sh and run it with:

sh scripts/test_M.sh

* TEST.WEIGHT is the path for the saved pytorch (.pth) model.

* There are four modes for TEST.TEST_MODE.

1. TEST.TEST_MODE 'test'
   • Use RRS[3] testing mode, which samples the first image of T snippets split from tracklet.
2. TEST.TEST_MODE 'test_0'
   • Sample first T images in tracklet.
3. TEST.TEST_MODE 'test_all_sampled'
   • Create N/T tracklets (all 1st image from T RRS snippets, all 2nd from T RRS snippets...), and average the N/T features.
4. TEST.TEST_MODE 'test_all_continuous'
   • Continuous smaple T frames, create N/T tracklets, and average the N/T features.

If you want to test on DukeV, you can just alter the corresponding arguments in scripts/test_M.sh.

New Evaluatoin Protocol

Change the TEST.NEW_EVAL False to TEST.NEW_EVAL True.

The details will be introduced soon.

Citation

@InProceedings{Liu_2021_CVPR,
    author    = {Liu, Chih-Ting and Chen, Jun-Cheng and Chen, Chu-Song and Chien, Shao-Yi},
    title     = {Video-Based Person Re-Identification Without Bells and Whistles},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
    month     = {June},
    year      = {2021},
    pages     = {1491-1500}
}

Reference

1. The structure of our code are based on reid-strong-baseline.
2. Some codes of our CF-AAN are based on axial-deeplab
3. Li, Shuang, et al. "Diversity regularized spatiotemporal attention for video-based person re-identification." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2018.

Contact

Chih-Ting Liu, Media IC & System Lab, National Taiwan University

E-mail : [email protected]