GitHub hosts a new PyTorch tool for recognizing soccer video highlights.
The repository ‘arturxe2/ASTRA’ showcases a PyTorch-based tool, which aligns with the research paper titled “ASTRA: An Action Spotting TRAnsformer for Soccer Videos”. This work, set to be featured at ACM MMSports’23, is a collaboration between Artur Xarles, Sergio Escalera, Thomas B. Moeslund, and Albert Clapés. Developers can adapt the code to both train and assess this innovative model using the SoccerNet-v2 collection, which comprises video and audio aspects.
For effective use, the software mandates that users integrate both Baidu visual features and audio patterns consistent with the SoccerNet framework. There are two primary scripts included: main.py for model training and main_challenge.py for evaluating performance in benchmarking challenges. These can be run with adjustable settings.
In addition, the reference format is available for those who wish to quote the research in their own work. Notably, some parts of the project’s design are a nod to the initial SoccerNet Action Spotting foundational package.
Read more: Github