The automated extraction of kinematic parameters from athletes in video footage allows for direct training feedback and continuous quantitative assessment of an athlete's performance. Recent developments in the field of deep learning enable the measurement of kinematic coefficients directly from human pose estimates. However, the detection quality decreases while errors and noise increase with the complexity of the scene. In aquatic training scenarios, for instance, continuous pose estimation suffers from several orthogonal errors like switched joint predictions between the left and right sides of the body. In this paper, we analyze different error modes and present a rectification pipeline for improving the pose predictions using merely joint coordinates. We show experimentally that joint rectification equally improves the detection of key-poses, which are essential for a continuous qualitative performance assessment and pose retrieval, as well as posture visualization for quantitative training feedback.
© Copyright 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE. Published by IEEE. All rights reserved.