Hightlight Video Detection in Figure Skating

Figure skating is an ornamental competitive sport with fancy technical moves. In particular, highlight videos in figure skating, which contain these elegant moves, have always been a favorite part of the vast audience. However, research in highlight video detection has not yielded much success. Previous researches mainly focus on detecting the falling action in the video rather than numerous technical moves. Therefore, we propose a segmentation method for the whole video to use Tube self-Attention for highlights detection. In particular, we have added a new module outside the existing network, which enables the editing and integration of the highlight moments of athletes in a single competition to produce highlight videos. Additionally, since few datasets have explored highlight actions in figure skating, we design a new dataset HS-FS (Highlight Shot in Figure Skating), which can be used to train the Tube Self-Attention model to satisfy highlight detection. Experiments show that the training accuracy obtained on the new dataset is 99.35% during training. Visualizations have demonstrated that our proposed methods could identify the highlight moment in figure skating videos.