Relative Pose Estimation for RGB-D Human Input Scans via Implicit Function Reconstruction

Abstract

To achieve a promising performance on relative pose estimation for RGB-D scans, a considerable overlap between two RGB-D inputs is often required for most existing methods. However, in many practical applications for human scans, we often have to estimate the relative poses under arbitrary overlaps, which is challenging for existing methods. To deal with this problem, this paper presents a novel end-to-end and coarse-to-fine optimization method. Our method is self-supervision which firstly combines implicit function reconstruction with differentiable render for RGB-D human input scans at arbitrary overlaps in relative pose estimation. The insight is to take advantage of the underlying human geometry prior as much as possible. First of all, for stable coarse poses, we utilize the implicit function reconstruction to dig out abundant hidden cues from unseen regions in the initialization module. To further refine the poses, the differentiable render is leveraged to establish a self-supervision mechanism in the optimization module, which is independent of standard pipelines for feature extracting and accurate correspondence matching. More importantly, our proposed method is flexible to be extended to multiview input scans. The results and evaluations demonstrate that our optimization module is robust for real-world noisy inputs, and our approach outperforms considerably than standard pipelines in non-overlapping setups.