Automated Scoring for Diving Events Using Action Quality Assessment

Sports like diving, gymnastics, figure skating, and skiing are dependent on professional judges to accurately score actions performed. The scores may be impacted by bias and human incompetence, especially at high levels, which can occasionally cause controversy. In the course of a competition, judges’ scores may be evaluated or occasionally updated using instant replay or recorded footage. During judging, most professional judges are looking for specific diving characteristics like body position (free, straight, tuck, and pike), the total number of “somersaults” and “twists” performed in the dive, the splash from the dive, the angle of entry into the pool, etc. All these details are necessary for determining how well the action of diving was performed by the diver. For this reason, extracting the right features from the video is very important for the AQA task. In this article, we suggest learning spatiotemporal aspects that explain two linked tasks, including calculating the AQA score and the fine-grained action identification. We demonstrate how our MTL method outperforms STL. With the aid of these models, we perform AQA score regression and have developed an interactive web application for the same.