Complementary geometric and optical information for match-propagation-based 3D reconstruction

Abstract

In this work, we consider the problem of propagation-based matching for 3D reconstruction, which deals with expanding a limited set of correspondences towards a quasi-dense map across two views. In general, propagation based methods capture well the scene structure. However, the recovered geometry often presents an overall choppy nature which can be attributed to matching errors and abrupt variations in the estimated local affine transformations. We propose to control the reconstructed geometry by means of a local patch fitting which corrects both the matching locations and affine transformations throughout the propagation process. In this way, matchings that propagate from geometrically consolidated locations bring coherence to both positions and affine transformations. Results of our approach are not only more visually appealing but also more accurate and complete as substantiated by results on standard benchmarks.