Kim Jeonghwan ; Gwon Mi-Gyeong ; Park Hyunwoo ; Kwon Hyukmin ; Um Gi-Mun ; Kim Wonjun

Abstract

This paper presents a simple yet powerful method for 3D human meshreconstruction from a single RGB image. Most recently, the non-localinteractions of the whole mesh vertices have been effectively estimated in thetransformer while the relationship between body parts also has begun to behandled via the graph model. Even though those approaches have shown theremarkable progress in 3D human mesh reconstruction, it is still difficult todirectly infer the relationship between features, which are encoded from the 2Dinput image, and 3D coordinates of each vertex. To resolve this problem, wepropose to design a simple feature sampling scheme. The key idea is to samplefeatures in the embedded space by following the guide of points, which areestimated as projection results of 3D mesh vertices (i.e., ground truth). Thishelps the model to concentrate more on vertex-relevant features in the 2Dspace, thus leading to the reconstruction of the natural human pose.Furthermore, we apply progressive attention masking to precisely estimate localinteractions between vertices even under severe occlusions. Experimentalresults on benchmark datasets show that the proposed method efficientlyimproves the performance of 3D human mesh reconstruction. The code and modelare publicly available at: https://github.com/DCVL-3D/PointHMR_release.