Master's Thesis : Action Spotting for Sport Video

Abstract

Action spotting is the challenging and promising task of locating temporal positions of specific action within untrimmed videos. Soccer is one of the most valuable application in Computer Vision and technological improvements are constantly encouraged. Current works do not tackle the question of detection methods practicality for real application in broadcast company. To bridge this information gap, one compares two state-of-the-art methods for action spotting : Context-Aware Loss Function Network (CALFNet) and Boundary-Matching Network (BMN). While CALFNet is designed for single frame spotting, BMN addresses the temporal proposal generation task. By developping and annotating a new kind of soccer dataset called LIVEX specially for this work, one studies, compares and conducts experiments on both methods and discuss them in terms of quantitative metrics and qualitative results, as video clips analysis. It is shown that CALFNet can be efficiently extended to larger number of actions and soccer videos for game understanding, while BMN achieves good qualitative performance on soccer dataset for practical application.