PC-HMR: Pose Calibration for 3D Human Mesh Recovery from 2D Images/Videos

The end-to-end Human Mesh Recovery (HMR) approach has been successfully used for 3D body reconstruction. However, most HMR-based frameworks reconstruct human by directly learning mesh parameters from images or videos, while lacking explicit guidance of pose in visual data. As a result, the generated often exhibits incorrect complex activities. To tackle this problem, we propose to exploit calibrate mesh. Specifically, develop two novel Pose Calibration frameworks, i.e., Serial PC-HMR and Parallel PC-HMR. By coupling advanced estimators HMR serial parallel manner, these can effectively correct with concise calibration module. Furthermore, since module is designed via non-rigid transformation, our flexibly bone length variations alleviate misplacement calibrated Finally, are based on generic complementary integration data-driven geometrical modeling. Via plug-and-play modules, they be efficiently adapted both image/video-based recovery. Additionally, have no requirement extra annotations testing phase, which releases inference difficulties practice. We perform extensive experiments popular benchmarks, Human3.6M, 3DPW SURREAL, where achieve SOTA results.