Date: August 2020.
Source: 2020 European Conference on Computer Vision (ECCV) Conference.
Abstract: We introduce BLSM, Our work aims at increasing the accuracy of data-driven rigged mesh representations. Our major contribution consists in revisiting the template synthesisan body mesh where bone scales are set prior to template synthesis, rather than the common, inverse practice. BLSM first sets bone lengths and joint angles to specify the skeleton, then specifies identity-specific surface variation, and finally bundles them together through linear blend skinning. We design these steps by constraining the joint angles to respect the kinematic constraints of the human body and by using accurate mesh convolution-based networks to capture identity-specific surface variation. We provide quantitative results on the problem of reconstructing a collection of 3D human scans, and show that we obtain improvements in reconstruction accuracy when comparing to a SMPL-type baseline. Our decoupled bone and shape representation also allows for out-of-box integration with standard graphics packages like Unity, facilitating full-body AR effects and image-driven character animation.
Article: BLSM: A Bone-Level Skinned Model of the Human Mesh.
Authors: Haoyang Wang, Riza Alp Guler, Iasonas Kokkinos, George Papandreou, and Stefanos Zafeiriou. Imperial College London, London, UK.