Self-consistency, Stereo, MDL, and Change detection

Abstract

Our main goal is to predict the accuracy of an output element of an existing stereo algorithm. Instead of relying extensively on ground truth, we apply independently the algorithm to subsets of images obtained by varying the camera geometry while the 3-D object geometry is kept constant. The self consistency methodology consists in collecting statistics in a scatter diagram of matches which should correspond to the same surface element in 3-D, along two dimensions in matching score in a normalized distance. We introduce a new matching score based on MDL theory, which is shown to be a better predictor of the quality of a Match than the traditional SSD score. The normalization is shown to make the statistics in variant to camera geometry. We demonstrate the potential of the methodology by two applications. In the first application, we compare stereo algorithms, matching scores, and window sizes. In the second application, we detect a change in shape between two sets of images taken at different times. We can discuss the application of self consistency to more general vision problems.

Keywords: Artificial Intelligence, Artificial Intelligence Center, AIC

Share this