Here you find the evaluation and training for TAAD on MultiSports Dataset. TAAD finished in the top spot in MultiSport challenge held in conjunction with ECCV 2022.
This repo heavily relies on PySlowFast so it contains a lot of stuff from there. It has the same License as PySlowFast, which is released under the Apache 2.0 license.
Multisport TAAD TCN model and multisports tracks generteted using Yolov5 and DeepSORT are availble to Download from Googl-Drive.
Please find installation instructions for PyTorch and PySlowFast in INSTALL.md.
Download Track from Googl-Drive. and Extract frame using multisport script provided by authors of Mulitsports.
Given the extracted frames and downloaed tracks and TAAD-TCN model from Googl-Drive. Now, you shoiuld be able to run TAAD-TCN which achives the best perfromance on Mulitsports.
Unfortunalty I do not have the bandwidth to reproduce this myself. But it should be doable with given current set of code base.
TAAD is written by Gurkirt Singh. I am happy to recieve a pull request for training reproduction.
If you find this useful in your research, please use the following BibTeX entry for citation.
@InProceedings{Singh_2023_WACV,
author = {Singh, Gurkirt and Choutas, Vasileios and Saha, Suman and Yu, Fisher and Van Gool, Luc},
title = {Spatio-Temporal Action Detection Under Large Motion},
booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
month = {January},
year = {2023},
pages = {6009-6018}
}Don't forget to cite orignal PySlowfast Repo.
@misc{fan2020pyslowfast,
author = {Haoqi Fan and Yanghao Li and Bo Xiong and Wan-Yen Lo and
Christoph Feichtenhofer},
title = {PySlowFast},
howpublished = {\url{https://github.com/facebookresearch/slowfast}},
year = {2020}
}