Residual Images Remove

Abstract

Real-world image sequences (e.g., recorded for vision-based driver assistance) are typically degraded by various types of noise, changes in lighting, out-of-focus lenses, differing exposures, and so forth. In past studies, illumination effects have been proven to cause the most common problems in correspondence algorithms. We address this problem using the concept of residuals, which is the difference between an image and a smoothed version of itself. In this paper, we conduct a study identifying that the residual images contain the important information in an image. We go on to show that they remove illumination artifacts using a mixture of synthetic and real-life images. This effect is highlighted more drastically when the illumination and exposure of the corresponding images is not the same.