Building Descriptors from Local Feature Neighborhoods for Applications in Semi-Dense 3D Reconstruction.

Abstract

This paper introduces a novel class of descriptors constructed solely from the extraction of a feature set's qualities of location, magnitude and orientation. Oriented Feature Constellations (OFC) descriptors are unlike typical feature descriptors that separate feature extraction from definition, where local underlying pixel information is used to form a unique high-dimensional vector for matching. We use other extracted features within the local neighborhood, or local constellation, to define unique descriptors for matching. This idea is demonstrated by creating a corner orientation extension to Features from Accelerated Segment Test (FAST) corners and use these features to create descriptors, called OFC-FAST. Using a standard two stage matcher to generate dense matches across test image sets, we compare our results with those obtained using Scale Invariant Feature Transform (SIFT) features. OFC-FAST descriptors were able to generate disparity maps of higher densities while using less computational time than the comparison method was able to achieve. We believe that the capability of OFC descriptors to quickly generate higher density disparity maps will be useful for applications in semi-dense 3D reconstruction.