Themain assumption of classical stereo matching is that matching pixels in stereo images have the same brightness values. However, this assumption is generally incorrect if we take into account the presence of noise in the images and the different illumination of the left and right images of the stereo-pair. In addition, pixel-by-pixel mapping does not work in areas where there is no real texture. Therefore, it is important to develop and use robust local features based on a certain neighborhood of the pixels being matched. In this study, we have developed and proposed to use a completely new local feature based on the Walsh-Hadamard transform. The transform in a local 8 x 8 window is easily encoded into a 64-bit value comparable in Hamming distance. The accuracy of such matching surpasses the use of a robust census feature that is close to the principle. Moreover, combining both features allows us to achieve the best stereo matching for methods that are not based on deep learning schemes.
展开▼
