Identifying Low Confidence Mesh Regions: Uncertainty Measures and Segmentation

3D digital models have become an important part of diverse applications ranging from computer games, virtual reality, architectural design to visual impact studies. One common method to create 3D models is to create a point cloud using laser scanners, structured lighting sensors, or image-based modelling techniques, and then construct a 3D mesh, and texture-map it using photographs of the observed scene. Attributed to the inherent properties of general 3D scenes such as occluded or inaccessible parts, reflective surfaces, lighting conditions or poor-quality inputs, 3D models produced by these approaches often exhibit unsatisfactory and erroneous mesh regions. In many cases, it is desirable to identify and extract such regions so that they can be constructed or corrected through other means. While much effort has been invested into the problem of 3D reconstructions, the task of evaluating existing models and preparing them for subsequent enhancement processes has been largely neglected. In this paper, we present a novel method for automatically detecting and segmenting mesh regions with low confidence in their correctness. The confidence estimation is achieved by exploiting and integrating various uncertainty measures such as geometric distances, normal variations and texture discrepancies. Low-confidence mesh regions are isolated and removed in such a way that the extracted region's boundary is as simple as possible in order to facilitate subsequent automatic or manual improvement of these regions. Segmentation is achieved by minimising an energy function that takes the genus and boundary length and smoothness of the extracted regions into account.