Occupancy Planes for Single-view RGB-D Human Reconstruction

  • 2022-08-04 18:59:56
  • Xiaoming Zhao, Yuan-Ting Hu, Zhongzheng Ren, Alexander G. Schwing
  • 3

Abstract

Single-view RGB-D human reconstruction with implicit functions is oftenformulated as per-point classification. Specifically, a set of 3D locationswithin the view-frustum of the camera are first projected independently ontothe image and a corresponding feature is subsequently extracted for each 3Dlocation. The feature of each 3D location is then used to classifyindependently whether the corresponding 3D point is inside or outside theobserved object. This procedure leads to sub-optimal results becausecorrelations between predictions for neighboring locations are only taken intoaccount implicitly via the extracted features. For more accurate results wepropose the occupancy planes (OPlanes) representation, which enables toformulate single-view RGB-D human reconstruction as occupancy prediction onplanes which slice through the camera's view frustum. Such a representationprovides more flexibility than voxel grids and enables to better leveragecorrelations than per-point classification. On the challenging S3D data weobserve a simple classifier based on the OPlanes representation to yieldcompelling results, especially in difficult situations with partial occlusionsdue to other objects and partial visibility, which haven't been addressed byprior work.

 

Quick Read (beta)

loading the full paper ...