Abstract
Disparity estimation is an ill-posed problem in computer vision. It is explored comprehensively due to its usefulness in many areas like 3D scene reconstruction, robot navigation, parts inspection, virtual reality and image-based rendering. In this paper, we propose a hybrid disparity generation algorithm which uses census based and segmentation based approaches. Census transform does not give good results in textureless areas, but is suitable for highly textured regions. While segment based stereo matching techniques gives good result in textureless regions. Coarse disparities obtained from census transform are combined with the region information extracted by mean shift segmentation method, so that a region matching can be applied by using affine transformation. Affine transformation is used to remove noise from each segment. Mean shift segmentation technique creates more than one segment of same object resulting into non-smoothness disparity. Region merging is applied to obtain refined smooth disparity map. Finally, multilateral filtering is applied on the disparity map estimated to preserve the information and to smooth the disparity map. The proposed algorithm generates good results compared to the classic census transform. Our proposed algorithm solves standard problems like occlusions, repetitive patterns, textureless regions, perspective distortion, specular reflection and noise. Experiments are performed on middlebury stereo test bed and the results demonstrate that the proposed algorithm achieves high accuracy, efficiency and robustness.
Highlights
Stereo vision is a fundamental problem in computer vision
The image pairs like Tsukuba, Teddy, Cones, Venus, and Sawtooth used for the evaluation purpose are popular and widely used by the stereo vision community
These stereo image pairs are well known for the combination of objects having different characteristics and are challenging for stereo matching
Summary
Stereo vision is a fundamental problem in computer vision. An extensive analysis of stereo matching algorithms. In regions having variable disparity different objects with relevant disparities will all be included in the census transform window which can offer a large amount of information to centre pixel by magnitude difference This clear difference in performance of census transform in texture and textureless regions is the motivation for our proposed approach. As adjacent pixels with similar colors have similar or continuous disparity, image segmentation is utilized to simplify the stereo problem [5]. Segment based stereo matching algorithm is used in our proposed approach along with census transform.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have