Enhanced point descriptors for dense stereo matching

We propose a novel local feature descriptor named Enhanced Point Descriptor (referred to as EPD) for dense stereo matching applications. The existing local feature descriptors, e.g., SIFT and SURF, can only be used to represent sparse image extreme points which make stereo matching sparsely. We design EPDs to represent common image points. To generate an EPD, we first build image characteristics vectors for neighborhood points around interest point in a specific sampled window. An EPD is a covariance matrix of characteristics vectors for all sampled points. The image characteristics we used to build vectors include HSV color, Gaussian-weighted gradient norms and orientations, which make EPD robust to rotation, perspective and illumination change. Experimental results show that EPD´s performance is superior to commonly used correlation windows methods in dense stereo matching.