Reducing Foot Artifacts with Physical Constraints
Yucheng Huang, Hong Yan · Frontiers in artificial intelligence and applications · 2025
Accurate prediction of foot position is a key element in estimating complex human poses. Recent advances aimed at improving the temporal smoothness of reconstructed motion trajectories and accurately predicting foot positions in most frames with only minor jitters. However, for infrequently occurring events, estimated foot positions may exhibit pronounced artifacts. In this paper, our objective is to mitigate these artifacts present in the current pose prediction model that uses monocular RGB videos. To achieve this, we introduce a dedicated temporal refinement network that is integrated with the current pose estimation framework. Our approach incorporates a fully connected layer network, which enforces a physical constraint to optimize foot positioning in video frames. We present a throughout evaluation of the suggested network and showcase the high-quality results achieved in reconstructing foot motion.