Official implementation of the paper ReFace: Improving Clothes-Changing Re-Identification With Face Features.
To evaluate the performance of our model, we provide a colab notebook. In this notebook, we first create an enriched gallery as described in the paper and then run the inference of our model using the enriched gallery.
In this paper we compare the results of our model on the LTCC, PRCC, and LaST datasets. The different datasets can be downloaded through the official pages of these datasets:
Inference on a custom dataset including person tracking, will be released soon, together with the 42Street dataset presented in the paper.
Our model relies on pre-trained face and ReID models and does not require any further training. See this folder for trained weights of the ReID model, trained by us on the original LTCC, PRCC, LaST and CCVID datasets (the checkpoints are automatically downloaded when running the colab notebook).
Below we provide the results achieved by our model on the clothes-changing settings in the different datasets.
| Dataset | PRCC | LTCC | LaST | CCVID |
|---|---|---|---|---|
| Top-1 | 83.7 | 74.8 | 75.8 | 89.2 |
| mAP | 66.7 | 48.4 | 29.6 | NaN |
In our work we use Simple-CCReID as the ReID module and Insightface as the face module. We thank them for their great works.
@article{arkushin2022reface,
title={ReFace: Improving Clothes-Changing Re-Identification With Face Features},
author={Arkushin, Daniel and Cohen, Bar and Peleg, Shmuel and Fried, Ohad},
journal={arXiv preprint arXiv:2211.13807},
year={2022}