Date: June 2014.
Source: 2014 IEEE Conference on Computer Vision and Pattern Recognition, Columbus, OH, USA, pp 3794-3801, doi: 10.1109/CVPR.2014.491.
Research Summary: New scanning technologies have increased the importance of 3D mesh data and the need for algorithms that can reliably align it. Surface registration underpins building full 3D models from partial scans, creating statistical shape models, shape retrieval, and tracking, and is especially challenging for non-rigid, articulated objects such as the human body. Existing synthetic datasets do not reflect the challenges present in real-world data, and establishing ground-truth correspondences for real 3D scans is difficult. FAUST addresses this with a mesh registration technique that combines 3D shape and appearance information to produce high-quality alignments, introducing a new dataset of 300 real-world scans of 10 people across a wide range of poses, along with an evaluation methodology. To achieve accurate registration, subjects were painted with high-frequency textures and validated through an extensive process to ensure accurate ground truth. Even so, the paper finds that current shape registration methods struggle with this real-world data, which underscores the difficulty of the problem that FAUST was built to address.
3dMD’s Role: FAUST’s 300 human body scans of 10 subjects in 30 poses each, were captured using a static 3dMDbody22 active stereophotogrammetry system composed of 22 Modular Camera Units (MCUs), each with a stereo camera pair, one or two speckle projectors, and a 5MP color camera. This is the predecessor dataset to Dynamic FAUST, which used a dynamic 60fps 3dMDbody22.u system to extend this same body of work into dynamic-3D/4D dense-surface motion capture.
Article: FAUST: Dataset and Evaluation for 3D Mesh Registration.
Authors: Federica Bogo, Javier Romero, Matthew Loper, Michael J Black. Max Planck Institute for Intelligent Systems, Perceiving Systems, Tuebingen, Germany.