Part-Aware Measurement for Robust Multi-View Multi-Human 3D Pose Estimation and Tracking

This paper introduces an approach for multi-human 3D pose estimation andtracking based on calibrated multi-view. The main challenge lies in finding thecross-view and temporal correspondences correctly even when several human poseestimations are noisy. Compare to previous solutions that construct 3D posesfrom multiple views, our approach takes advantage of temporal consistency tomatch the 2D poses estimated with previously constructed 3D skeletons in everyview. Therefore cross-view and temporal associations are accomplishedsimultaneously. Since the performance suffers from mistaken association andnoisy predictions, we design two strategies for aiming better correspondencesand 3D reconstruction. Specifically, we propose a part-aware measurement for2D-3D association and a filter that can cope with 2D outliers duringreconstruction. Our approach is efficient and effective comparing tostate-of-the-art methods; it achieves competitive results on two benchmarks:96.8% on Campus and 97.4% on Shelf. Moreover, we extends the length of Campusevaluation frames to be more challenging and our proposal also reachwell-performed result.