READ: Reciprocal Attention Discriminator for Image-to-Video Re-Identification

Minho Shim, Hsuan-I Ho, Jinhyung Kim, and Dongyoon Wee

This repository contains demo software for the READ (ECCV 2020).

Environment

CUDA 9.0
Python 3.5.2
PyTorch 0.4.1
OpenCV 3.4.7
H5py
SciPy

Dataset Preparation (MARS)

Unzip bbox_train.zip and bbox_test.zip, alongside with mat/txt files. (Links: zips, mats/txts from the MARS authors' repo)

The directory will look like:

path/to/mars_raw/
|-- bbox_train/
|-- bbox_test/
|-- info/

Then run:

python dataset/mars/mars_jpgs_to_h5.py --top-dir path/to/mars_raw/ --output-dir path/to/mars_h5/
python dataset/mars/mars_extract_json.py -d path/to/mars_h5/ -o path/to/mars_h5/mars.min.json

Download weights

Download links to weight files for MARS, and DukeMTMC-VideoReID.

Run (MARS)

python play.py --test --dataset mars --json-path path/to/mars_h5/mars.min.json --h5-dir path/to/mars_h5/ --checkpoint-files '{"all":"[downloaded_checkpoint_path]"}'

By default, it uses all available GPUs in the machine. It could be limited with CUDA_VISIBLE_DEVICES, e.g. CUDA_VISIBLE_DEVICES=0,1,2,3 python ...

Adjust --batch-size accordingly to your GPU capacity. In our case, P40 or V100 GPUs are used with the default batch size of 32.

Sample Results (MARS)

python play.py --test --dataset mars --checkpoint-files '{"all":"/volume/READ/mars_demo.tar"}' --json-path /volume/dataset/mars_h5/mars.min.json --h5-dir /volume/dataset/mars_h5/

...

Test Results:
mean_average_precision: 0.7039292454719543
mean_top10: 0.9383838176727295
mean_top5: 0.9207070469856262
mean_top1: 0.8146464824676514

Citation

@InProceedings{shim2020read,
author = {Shim, Minho and Ho, Hsuan-I and Kim, Jinhyung and Wee, Dongyoon},
title = {{READ}: {R}eciprocal Attention Discriminator for Image-to-Video Re-Identification},
booktitle = {European Conference on Computer Vision (ECCV)},
year = {2020}
}

Lincense and References

This software is for non-commercial use only. Imported/modified codes contain their own references inside each code.

minostauros / READ