Selecting Surface Features for Accurate Multi-Camera Surface Reconstruction

TL;DR: A novel feature detector for selecting local textures that are suitable for accurate multi-camera surface reconstruction, and in particular planar patch fitting techniques, is proposed.

Abstract: This paper proposes a novel feature detector for selecting local textures that are suitable for accurate multi-camera surface reconstruction, and in particular planar patch fitting techniques. This approach is in contrast to conventional feature detectors, which focus on repeatability under scale and affine transformations rather than suitability for multi-camera reconstruction techniques. The proposed detector selects local textures that are sensitive to affine transformations, which is a fundamental requirement for accurate patch fitting. The proposed detector is evaluated against the SIFT detector on a synthetic dataset and the fitted patches are compared against ground truth. The experiments show that patches originating from the proposed detector are fitted more accurately to the visible surfaces than those originating from SIFT keypoints. In addition, the detector is evaluated on a performance capture studio dataset to show the real-world application of the proposed detector.