Awesome

CLIP-ReIdent

[Paper] [Checkpoints] [Challenge]

Code for the paper: CLIP-ReIdent: Contrastive Training for Player Re-Identification

Approach

Reformulation of the contrastive language-to-image pre-training approach of CLIP to a contrastive image-to-image training approach using the InfoNCE loss as training objective.

Usage

Steps for Training and Evaluation:

1. get data: download_data.py
2. create DataFrames: preprocess_data.py
3. training: train.py
4. evaluation: evaluate.py
5. final predictions: predict.py

All settings are done by the configuration dataclass at the beginning of the scripts.

download_data.py downloads and unzips the challenge data from the provided challenge toolkit.

The following data structure is used for all scripts:

clip_reid
├── data/
│   ├── data_reid/
│   │   ├── reid_challenge/
│   │   │   ├──gallery/
│   │   │   └──query/
│   │   ├── reid_test/
│   │   │   ├──gallery/
│   │   │   └──query/
│   │   ├── reid_training/
│   │   ├── challenge_df.csv
│   │   └── train_df.csv
│   └── synergyreid_data.zip
│
├── clipreid/
│   ├── dataset.py
│   ├── evaluator.py
│   ├── loss.py
│   ├── metrics.py
│   ├── model.py
│   ├── rerank.py
│   ├── trainer.py
│   ├── transforms.py
│   └── utils.py
│
├── model/
│   └── ...checkpoints...	
│
├──  download_data.py
├──  evaluate.py
├──  predict.py
├──  preprocess_data.py
└──  train.py

If you find this repository useful, please consider citing:

@inproceedings{habel2022clipreident,
  title={CLIP-ReIdent: Contrastive Training for Player Re-Identification},
  author={Habel, Konrad and Deuser, Fabian and Oswald, Norbert},
  booktitle={Proceedings of the 5th International ACM Workshop on Multimedia Content Analysis in Sports (MMSports’22), October 10, 2022, Lisboa, Portugal},
  year={2022}
}