Learning salient elements from videos using structural and temporal information

Resumo

This project proposes a novel framework to learn important elements from sports video sequences. The basic idea behind the proposal is that some key elements are salient in sports video and we would like to devise strategies for automatic learning them. For instance, in table tennis, despite the large possible variations in terms of camera pose, background, colors, etc, some key elements are consistently present in such videos: the table, two players moving around it, the rackets, the ball. The authors of this proposal are currently co-supervising a PhD project (Henrique Morimitsu, proc. FAPESP DR 12/09741-0 and BEPE 13/08258-7) which develops a novel framework for table and player detection in table tennis matches videos, which combines low-level features, temporal and structural information. The goal of this project submitted to CNRS/FAPESP is to go one step further by learning automatically the main elements in an unsupervised manner from video samples. Our idea is to use this proposal to recrute a Post-Doctoral researcher to develop the project (the PD scholarship will be submitted in due time in case this project is supported). (AU)