Deep dynamic feature alignment-based few-shot learning for figure skating action localization
Mingyang Wang, Yiqun Pang, Qiurui Wang, Dan Chen · 2023
Temporal action localization is a challenging and important task in computer vision. Although great progress has been made in recent years in this area with the help of deep learning, few jobs focus on applications on competitive sports, such as figure skating. To better analyze the players’ actions in figure skating, action localization is needed. Hence we collect 82 videos from worldwide figure skating contests, including championship and grand prix, etc. These videos are annotated by experts for action start timestamps, action end timestamps, and action names. We further analyze these videos by our proposed Deep Dynamic Feature Alignment (DDFA) method, which employs deep convolutional features and learns the action pattern between the target action and reference action. Experiments on models’ performances and results based on these datasets are presented, showing the proposed methods can predict proper action candidates for raw figure skating videos.