Date: January 2023.
Source: 17th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2023)
Research Summary: A decade after Headspace was first captured, the same core research team returned to the dataset to introduce Laplacian ICP (L-ICP), a new, highly efficient method for registering the FaceWarehouse head template to raw 3D head scans, a foundational step in building any statistical shape model. Standard non-rigid registration methods, such as N-ICP (Non-rigid Iterative Closest Points), are computationally expensive; L-ICP uses Laplace-Beltrami regularization and Mutual Nearest Neighbour (MNN) correspondence matching within a coarse-to-fine, staged framework, guided by facial landmarks, ear landmarks, and an automatically extracted symmetry contour, achieving comparable registration quality roughly 26 times faster than the Per-Vertex Affine Constraint (PVAC) approach used in standard N-ICP. Evaluated on the real-world 3dMD Headspace dataset, the paper also introduces a new benchmark for 3D non-rigid registration generally, based on transferring manual contour annotations from raw scans onto the registered template and measuring how consistently different subjects’ annotations land on the same template location.
3dMD’s Role: Evaluated on the real-world 3dMD Headspace dataset, using 675 subjects for the main annotation-transfer benchmark and a 118-subject subset for processing-time comparisons.
Article: Laplacian ICP for Progressive Registration of 3D Human Head Meshes.
Authors: Nick Pears, William Smith, Hao Sun. Department of Computer Science, University of York. Hang Dai, Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), UAE.