Performance Analysis of Feature Point Detectors in SFF-Inspired Relative Depth Estimation

This paper is a part of the research that attempts to develop a new method for estimating relative depth of scene based on focus cue inspired by the Shape From Focus (SFF) technique which basically is a 3-D vision technique aiming at scene reconstruction. The proposed method is developed for scene with objects whose geometrical dimensions are of macro-level unlike the conventional SFF algorithms which could deal with extremely small objects in order to avoid parallax. This is essentially because SFF techniques involve motion of either the camera or the object causing structure dependent pixel motion in the image. This hinders finding the corresponding pixels in the sequence of images to perform a measure of image focus and resulting in erroneous reconstruction. The parallax effect is tackled using a matching technique which tracks point correspondences in the image sequence. The points with good local contrast are extracted using the so-called interest point detectors. The paper analyses five different point detectors used to extract feature points in the image. The point detectors were analysed for a number of parameters like repeatability, false detects, matchability and information content. The paper compares their performance and attempts to bring out their advantages and limitations. Three different conditions viz., uniform illumination, non-linear illumination changes, presence of impulse noise in the image were considered. The image processing method is validated for two different textures, one being a repetitive pattern and other a non-repetitive pattern.