Visibility Analysis for New View Synthesis from Images

This work introduces a new analysis procedure to measure availability of information to render a new view, given a set of views from a camera in linear motion, or a linear array of cameras. The given set can be interpolated to generate a new view using a crossed-slit camera model. The analysis consists of finding regions in the neighborhood of the linear camera positions, where a full view, partial view, or no view can be generated. A visibility parameter, V, is introduced to quantitatively measure the amount of information that can be interpolated. The present work differs from previous methods which aim at minimizing the number of input viewpoints while ensuring full coverage of a specific object. Our work analyzes the visibility for a set of views captured by a linear camera array, and can be extended to a set of images from a 2D camera mesh. The result of this analysis, the visibility, V, is computed and verified with respect to the interpolated views. Visibility analysis helps in identifying blind areas while generating videos for a new camera trajectory. It is also useful in planning best scene capturing strategy for the purpose of generating new views, and finds application in virtual walkthroughs, special effects, the free viewpoint television, remote education, etc.