Stereo matching using normalized cross-correlation in LogRGB space
Gaojian Li · 2012
Stereo matching is the most important branch in computer vision. Due to various matching primitives, the stereo matching approaches are divided into three categories, namely, area-based matching method, feature-based matching method and phase-based matching method. In majority of existing matching algorithms, it is a critical assumption that the corresponding color values of stereo images are similar to each other. However, various radiometric factors affect image color in practice, such as lighting geometry, illuminant color, and imaging devices. Therefore, the conventional stereo matching methods based on such assumption may not obtain the expected results. In this paper, we present a new stereo matching measure that is unaffected by radiometric variations between stereo images. We use the color formation model based on Retinex [1] theory in logRGB space [2], and derive the disparity between two images using normalized cross-correlation. The occluded pixels are filtered out by cross-checking[3]. Experiments show that our method is robust to various radiometric factors between stereo images.