Comparative Analysis of state-of-art pre-trained Human Pose Estimation models in underwater condition

Human Pose Estimation (HPE) is essential in computer vision, with applications in sports, assisted living, and gaming. Pre-trained HPE state-of-the-art models like OpenPose, MoveNet and MediaPipe have strengths and weaknesses that have yet to be discovered (advantages and disadvantages are not completely known yet). While it is true that applying HPE underwater to predict swimmers' movements is not a very explored area due to the difficulties this environment represents, it could bring numerous benefits, especially in the fields of sport and human body correction. The study proposes an analysis of pre-trained models (all layers frozen) performance by evaluating their accuracy predictions over 3 videos of professional swimmers doing a dolphin kick underwater using 3 different processes to see which one fits better to each model: without pre or post-processing (Type 1), with pre-processing (Type 2) and with pre and post-processing (Type 3). Results concluded that MediaPipe with Type 3 processing and a confidence of 50% was the most effective for underwater HPE, while OpenPose and MoveNet did not perform well in these conditions.