RemoCap: Disentangled Representation Learning for Motion Capture

Reconstructing 3D human bodies from realistic motion sequences remains achallenge due to pervasive and complex occlusions. Current methods struggle tocapture the dynamics of occluded body parts, leading to model penetration anddistorted motion. RemoCap leverages Spatial Disentanglement (SD) and MotionDisentanglement (MD) to overcome these limitations. SD addresses occlusioninterference between the target human body and surrounding objects. It achievesthis by disentangling target features along the dimension axis. By aligningfeatures based on their spatial positions in each dimension, SD isolates thetarget object's response within a global window, enabling accurate capturedespite occlusions. The MD module employs a channel-wise temporal shufflingstrategy to simulate diverse scene dynamics. This process effectivelydisentangles motion features, allowing RemoCap to reconstruct occluded partswith greater fidelity. Furthermore, this paper introduces a sequence velocityloss that promotes temporal coherence. This loss constrains inter-framevelocity errors, ensuring the predicted motion exhibits realistic consistency.Extensive comparisons with state-of-the-art (SOTA) methods on benchmarkdatasets demonstrate RemoCap's superior performance in 3D human bodyreconstruction. On the 3DPW dataset, RemoCap surpasses all competitors,achieving the best results in MPVPE (81.9), MPJPE (72.7), and PA-MPJPE (44.1)metrics. Codes are available at https://wanghongsheng01.github.io/RemoCap/.