Soares João V. B. ; Shah Avijit

Abstract

This brief technical report describes our submission to the Action SpottingSoccerNet Challenge 2022. The challenge was part of the CVPR 2022 ActivityNetWorkshop. Our submission was based on a recently proposed method which focuseson increasing temporal precision via a densely sampled set of detectionanchors. Due to its emphasis on temporal precision, this approach had shownsignificant improvements in the tight average-mAP metric. Tight average-mAP wasused as the evaluation criterion for the challenge, and is defined using smalltemporal evaluation tolerances, thus being more sensitive to small temporalerrors. In order to further improve results, here we introduce small changes inthe pre- and post-processing steps, and also combine different input featuretypes via late fusion. These changes brought improvements that helped usachieve the first place in the challenge and also led to a new state-of-the-arton SoccerNet's test set when using the dataset's standard experimentalprotocol. This report briefly reviews the action spotting method based on densedetection anchors, then focuses on the modifications introduced for thechallenge. We also describe the experimental protocols and training procedureswe used, and finally present our results.