Evaluation of feature detection algorithms for structure from motion

Abstract

Structure from motion is a widely-used technique in computer vision to perform 3D reconstruction. The 3D structure is recovered by analysing the motion of an object, based on its features, over time. The typical steps involved in SFM are feature detection, feature matching and determining the motion and pose of the cameras. For each step, a number of different algorithms may be used. Little research has however been done into the effectiveness of the different feature detection algorithms such as Harris corner detectors and feature descriptors such as SIFT (Scale Invariant Feature Transform) and SURF (Speeded Up Robust Features) given a set of input images. This paper implements state-of-the art feature detection algorithms and evaluates their results on a given set of input images. The evaluation will be preformed by comparing the calibration data, the fundamental matrix and the rotation and translation errors extracted from each algorithm with ground truth data.