Multi-modal Scene Reconstruction using Perceptual Grouping Constraints

In this work we propose a scheme integrating perceptual grouping into stereopsis to reduce the ambiguity of those early processes. We propose a simple perceptual grouping algorithm that - in addition to the geometric information makes use of a novel multi-modal affinity measure between local primitives. We then use this group information to 1) disambiguate the stereopsis by enforcing that stereo matches preserve groups; and 2) correct the reconstruction error due to the image pixel sampling using a linear interpolation over the groups. We show quantitative and qualitative demonstrations of those processes on a variety of sequences.