Date: August 2020.
Source: 2020 European Conference on Computer Vision (ECCV) Conference.
Research Summary: BLSM improves the accuracy of rigged 3D body meshes, the standard representation used to animate human characters, by changing the order in which the body is built. Most existing models, including SMPL, generate the body’s shape first and then work out the skeleton afterward, fitting joints to a mesh that already exists. BLSM flips this: it sets bone lengths and joint angles first to build the skeleton, then adds identity-specific shape detail on top, and finally combines both using linear blend skinning (LBS), the standard technique for deforming a mesh around a skeleton. Joint angles are constrained to match the way human joints actually move, and shape detail is captured using mesh convolution-based networks, a more powerful alternative to the simpler linear methods older models rely on. The result outperforms an SMPL-type baseline at reconstructing real-world 3D human scans, and because bone and shape are kept separate, the model plugs directly into standard graphics software like Unity for full-body AR effects and character animation.
3dMD’s Role: BLSM’s skinning weights, the values that determine how strongly each part of the mesh follows each bone, are trained directly on D-FAUST, a 3dMD-captured motion dataset. Most of D-FAUST’s subjects are used for this training, with two held back to test how well the model generalizes. Beyond D-FAUST, BLSM is evaluated a second time on an entirely separate in-house 3dMDbody dataset, 20 subjects captured as 4D sequences at 30fps, spanning a range of body types and poses. The model’s shape component, by contrast, is trained on CAESAR, a 1998 to 2000 U.S. Air Force-led laser-scan anthropometric survey of 3 poses, built for industrial sizing rather than 3D/4D motion research.
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.