Next: Geometric Model Construction Up: Computer Vision Previous: Object Description and

Curvature Consistency in Surface Reconstruction

Authors: [tex2html_wrap4150]S. Mathur, G. Soucy, F.P. Ferrie

Investigator username: ferrie

Category: perception

Subcategory: computer vision

Data acquired from sensors such as television cameras and laser rangefinders provide only indirect information about the scene being imaged. As a result the data are often ambiguous as well as being subject to the usual noise and quantization error. The primary goal of reconstruction is to reduce data to stable descriptions that can be used in a robust manner by different processes operating on it. Because our interests are related to characterizing the geometric structure of objects, we are interested in the application of reconstruction processes to produce stable surface descriptions. The approach that we are investigating is based on a class of algorithms developed by Zucker and his co-workers that iteratively minimize a residual form related to the satisfaction of local constraints on the curvature of the surface.

Our interest is in how these methods can be applied and extended to various problems in surface reconstruction, e.g. shape-from-X, both in the spatial and temporal domains. Previous work demonstrated how the curvature consistency framework could be used in the reconstruction of range images and as a second stage of processing to improve the results of local shading analysis. Recent efforts have dealt with extending the framework to include multiple sources of information (sensor fusion) as well as better preserving the local structure of surface discontinuities.


baron@cim.mcgill.ca