ADAPTIVE GABOR FILTERS FOR PHASE-BASED DISPARITY ESTIMATION

Abstract

The phase-difference-based technique has become a widespread method for depth and optical flow estimation because of superior performance and better theoretical groundings. The technique is based on the convolution of the stereo image pair with Gabor filters. Gabor filters contains two parameters, the width and the tuning frequency. In order to optimize performance, these parameters have to be chosen in accordance to the characteristics of the visual signal. In this article we propose an automatic technique to locally adapt the filter parameters to the input signal. In the first part, we analyze the performance of the phase-difference-based technique for disparity estimation with respect to the choice of the Gabor filter parameters. In particular we characterize the effects of phase nonlinearities on the quality of disparity estimates. In the second part, a novel technique is introduced that reduces phase nonlinearity by means of an adaptive mechanism for the tuning frequency. The performance improvement produced by the adaptive filter is demonstrated using different types of images. Results show the proposed technique allows a significant improvement of disparity estimation.