|M.Sc Student||Marciano Abraham|
|Subject||Design and Implementation of a Stereo|
Algorithm Robust to Texture-Less Scenes
|Department||Department of Electrical Engineering||Supervisor||Professor Guy Gilboa|
|Full Thesis text|
In recent years, stereo matching has remained a prominent research area in computer vision. Yet, many of the developed algorithms are parameter-dependent and assessed on classical highly textured benchmarks that are not necessarily representative of typical indoor scenes. Our work consisted of finding an intuitive, fast and robust model that can also be applied to texture-less scenes. In contrast with the latest models, we resort to computationally simple methods that can potentially yield fast results.
In our study, segmentation is used as a preprocessing step in order to avoid many of the problems inherent to stereo correspondence. The algorithm is based on L0 gradient minimization combined with Mean-Shift clustering to get segmented stereo images. Super-pixels are obtained and then merged to give an approximate object-level segmentation. This is followed by a stereo segment-pair matching based on color and geometrical features. The algorithm is tested on our own rectified stereo pictures obtained after camera calibration. A summary of the major current directions in stereo is also given, along with experimentations of techniques based on disparity estimation at edges and anisotropic diffusion