This is the official github repository for the Vidpress-Sports project in Baidu Research.
Baidu soccer embedding features for 550 complete soccer game videos in SoccerNetv2
We release our soccer embedding features here. 89GB storage space is needed to download the zipped features.
The features can be used to reproduce our results in the CVPR 2021 ActivityNet Challange, Temporal Localization track, SoccerNet Challenge. We are ranked No.1 in both action spotting task and replay grounding task.
The released embedding features are the concatenation of five features extracted using five finetuned video action recognition models. For details, please refer to our tech report on arXiv. The features can also be used to support other soccer video understanding research.
We now have a fully open-sourced inference and training pipeline here. I recommend the new features pipeline since one can do inference on matches not included with the 550 games (feel free to share and request there) and you can do more tuning / training with it. Without extra tuning and augmentation on the stage 2 action spotting training, these features get a_mAP visibility all: 0.695263465366132 which is very close to the 5 model ensemble we share in this repo.
If you find our features or tech report helpful, please cite:
@misc{zhoukangcheng2021,
title={Feature Combination Meets Attention: Baidu Soccer Embeddings and Transformer based Temporal Detection},
author={Xin Zhou and Le Kang and Zhiyu Cheng and Bo He and Jingyu Xin},
year={2021},
eprint={2106.14447},
archivePrefix={arXiv},
primaryClass={cs.CV}
}