Abstract
Stereo matching is one of the main topics in computer vision. It consists in finding in two images of a same scene, taken from different viewpoints, the pairs of pixels which are the projections of a same scene point. Since the last twenty years, many local and global methods have been proposed to solve this problem. More recently, according to a reference evaluation protocol in the community, region-based methods showed interesting result in small-baseline binocular stereo (where images are taken nearby). The idea is to apply a colour segmentation algorithm on the images assuming that each pixel within a segment belongs to a same object surface. Then, the parameters of a surface model are computed, in the disparity space, for each segment according to initial disparities usually computed with a local method. Finally, a global optimization is performed to refine the results.
A contribution of this thesis deals with a special kind of local method called seeds propagation. The search area of a correspondent is reduced to the neighbourhoods of reliable matches called seeds. This can help to reduce the computation time and to avoid some ambiguities. However, the success of such a method depends on the choice of these seeds. In this dissertation, we give a study of the seeds selection step. We focus on feature points matching. These are special points in the image with interesting characteristics for a given application. In our case, we need pixels that can be matched with high confidence. We compare fourteen different well-known detectors linked to five correlation measures. Some of these measures are meant to be robust to one of the main challenge in stereo matching: depth discontinuities. Besides, this study gives advice on the choice of the parameters of the different techniques to be able to find the best solutions according some given criteria. These parameters are estimated using machine learning. Then, these seeds are used with two approaches of propagation and the results are evaluated.
Another contribution deals with a new region-based approach for dense stereo matching. Different colour segmentations are used. Then, many instances of a surface model are computed for the different regions according to initial disparities selected randomly. For each pixel, each instance gives a disparity value regarded as a vote. Finally, the most voted value is selected as the final disparity. This approach is relatively easy to implement and very effective giving competitive results among the state of the art.
