Decomposed Key-Point Detector for Swimming Pool Localization

Pool localization is an essential prerequisite for swimmer analysis or performance measurement. Automated analysis studies on swimming pools have often been proposed due to the increase in broadcast videos and the development of machine learning techniques, but there have been few studies dealing with GPU memory usage and latency in terms of practical analysis. This work proposes an efficient swimming pool key-point detection method utilizing a U-Net-based decomposed detector, for robust detection regardless of camera parameters. Experiments show the proposed detector has higher accuracy than the original model, while reducing latency and memory usage by 4.80 and 2.91 times, respectively.