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.
2021-06-13
:
2021-06-18
:
The video tracklet will be re-Detected, linked (tracking) and padded to the original image size, as follow.
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/
We use Python 3.6, Pytorch 1.5 and Pytorch-ignite in this project. To install required modules, run:
pip3 install -r requirements.txt
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.
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
You can use the scripts scripts/AA_D.sh
, scripts/NL_D.sh
, & scripts/baseline_D.sh
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
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
.
TEST.TEST_MODE 'test'
TEST.TEST_MODE 'test_0'
TEST.TEST_MODE 'test_all_sampled'
TEST.TEST_MODE 'test_all_continuous'
If you want to test on DukeV, you can just alter the corresponding arguments in scripts/test_M.sh
.
Change the TEST.NEW_EVAL False
to TEST.NEW_EVAL True
.
The details will be introduced soon.
@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}
}
Chih-Ting Liu, Media IC & System Lab, National Taiwan University
E-mail : jackieliu@media.ee.ntu.edu.tw