MLA-LSTM: A Local and Global Location Attention LSTM Learning Model for Scoring Figure Skating

Video-based scoring using neural networks is a very important means for evaluating many sports, especially figure skating. Although methods action quality have been proposed, there no uniform conclusion on the best feature extractor and clip length existing methods. Furthermore, during aggregation stage, these cannot accurately locate target information. To address tasks, firstly, we systematically compare effects of skating model with three different extractors (C3D, I3D, R3D) four segment lengths (5, 8, 16, 32). Secondly, propose Multi-Scale Location Attention Module (MS-LAM) to capture location information athletes in video frames. Finally, present novel Multi-scale Attentive Long Short-Term Memory (MLA-LSTM), which can efficiently learn local global sequence each video. In addition, our proposed has validated Fis-V MIT-Skate datasets. The experimental results show that I3D 32 frames per second are tasks. outperforms current state-of-the-art method hybrid dynAmic-statiC conText-aware attentION NETwork (ACTION-NET), (by 0.069 Spearman’s rank correlation). it achieves average improvements 0.059 compared convolutional skip Self-attentive LSTM (MS-LSTM). It demonstrates effectiveness models learning score videos.