Hong-Tan based ICP registration for partially overlapping range images
Proceedings of the 10th International Conference on Virtual Reality Continuum and Its Applications in Industry
| Hi-index | 0.00 |
The world around us consists of objects of vastly varying shapes, sizes, and geometric complexity. We propose algorithms for capturing, comparing, and analyzing such 3D geometry. In the first part, we automate various stages of a standard shape acquisition pipeline. Typically, a 3D scanner captures object geometry from multiple directions. Each viewpoint gives a partial geometric model or a scan. Scan registration involves stitching these scans together to form one consolidated model. We present an algorithm for the automatic global rigid alignment of two 3D shapes, without any assumptions about their initial poses. We implicitly evaluate all the possible correspondence assignments between a set of selected feature points without having to explicitly enumerate each of the correspondences. The rough initial alignment is further locally refined using a minimization of the squared distance between the underlying surfaces. We locally approximate the squared distance field using quadratic functions and then develop a linear system to solve for the local aligning rigid transform. Model registration often results in an incomplete representation of the scanned object. We present a novel approach for plausible 3D model completion using geometric priors. Our method retrieves suitable context models from a model database, warps them to conform with the input data, and consistently blends the warped models to obtain the final consolidated 3D shape. In the second part, we introduce two shape analysis tools. We present an algorithm that processes geometric models and efficiently discovers and extracts a compact representation of their Euclidean symmetries. These symmetries can be partial, approximate, or both. The extracted symmetry graph representation captures important high-level information about the structure of the model. We also propose a compact shape signature based on probabilistic fingerprints. Our method is robust to noise, invariant to rigid transforms, handles articulated deformations, and effectively detects partial matches. These compact fingerprints are used to efficiently estimate similarity across multiple 3D shapes where directly evaluating similarity is expensive and impractical. We demonstrate the utility of all our algorithms for a wide variety of geometry processing applications on a range of scanned geometric models of varying sizes, complexity, and details.