In this work, we propose a method to detect swimmers in unconstrained swimming videos, using a Unet-based model trained on a small dataset. Our main motivation is to make the method accessible without spending much time or money in annotation or computation while maintaining performances. The swimming videos can be recorded from various locations with different settings (distance and angle to the pool, light conditions, reflections, camera resolution), which alleviates a lot of the usual video capture constraints. As a result, our model reaches top-performances in detection compared to other methods. Every algorithm described in the paper is accessible online at https://github.com/njacquelin/swimmers_detection.
© Copyright 2022 Machine Learning and Data Mining for Sports Analytics. MLSA 2021. Communications in Computer and Information Science. Veröffentlicht von Springer. Alle Rechte vorbehalten.