ReFace: Improving Clothes-Changing Re-Identification With Face Features

Official implementation of the paper ReFace: Improving Clothes-Changing Re-Identification With Face Features.

Quick start

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.

Usage Example - LTCC

Gallery Enrichment

ltcc --dataset_path <path-to-dataset> --detection_threshold 0.8 --similarity_threshold 0.5 --device <device> --CC

Inference

ltcc AIM --reid_config <path-to-reid-config-file> --dataset_path <path-to-dataset> --detection_threshold 0.7 --similarity_threshold 0.5 --alpha 0.75 --reid_checkpoint <path-to-checkpoints> --device <device>

Notes

• To download the datasets see the original pages of each dataset (listed below).
• The reid_config files of the supported models can be found under ReIDModules\<reid-model>\configs.
• Checkpoints of the different models can be downloaded from here.

Datasets

Existing Benchmarks

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:

• CCVID
• LTCC
• PRCC
• LaST
• VC-Clothes

The 42street Dataset

The 42street dataset can be downloaded from the following link:

• dataset
• extra-data-1
• extra-data-2

Dataset Structure

• gallery: folder with annotated crops - 16,668 images of 13 identities + 1 category for unidentified persons.
• test:
  • vids: raw videos that were used for testing - 10 videos of ~17 seconds each.
  • tracklets: annotated tracklets from the test videos - 26,427 images in 239 tracklets.
• extra-data: a folder with unannotated crops taken from the same part in the play as the test videos (downloaded separately).

To use the extra data, download both folders above and extract them to the extra-data folder.

Trained model weights

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).

Results

Below we provide the results achieved by our model on the clothes-changing settings in the different datasets.

Dataset	PRCC	LTCC	LaST	VC-Clothes	CCVID	42Street	42Street (w. extra-data)
Top-1	81.9	76.3	78.0	94.9	89.2	75.0	92.2
mAP	58.8	42.3	37.2	88.9	NaN	NaN	NaN

Acknowledgments

In our work we use Simple-CCReID as the ReID module and Insightface as the face module. We thank them for their great works.

Citation

@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}