The Feature Detection on the Homogeneous Surfaces with Projected Pattern

Abstract

In this article we deal with one of the fundamental problems in the area of the 3D reconstruction for objects with homogeneous surface such as, inter alia, human body or sculptures. The interest point detection on typical photos with many differing elements and changing intensities is already well-solved issue. Considerable difficulty and novelty is the interest point detection for homogeneous surfaces. To reconstruct such surfaces from images we have to artificially produce as many elements on surface as needed to allow proceed with the 3D coordinate’s extraction process with desired density. Four methods were selected. The first, definitely the best documented was the Harris corner detector. Next was the Nobel’s version of auto-correlation, the other was the minimum eigenvalue method known as the Kanade-Tomasi algorithm and the last tested method was the fast radial feature detector known as the Loy-Zelinsky algorithm. Chosen methods are well-known on the 3D reconstruction theatre, well implemented and documented, efficient in the terms of computational complexity. Also some image enhancements were utilized before feature extraction to improve the detection process. It was shown that the best choice was the Nobel’s version of auto-correlation function and a very interesting candidate for further research is the Loy-Zelinsky method.