Feature based dense disparity estimation

Dense disparity estimation for stereo imaging is an important building block for several computer vision applications. Feature based disparity estimation technique provides disparity at only few of the known feature points and thus does not give a good estimate of the complete region. In this paper, a novel framework of dense disparity estimation using feature correspondences are proposed, which employs image segmentation and feature based segment matching technique as the core methodology. It avoids any optimization technique and is thus not an NP-hard problem. The initial features obtained using SURF detector is used for matching segments indirectly. Segments which do not contain any feature point are matched via epipolar constraint by randomly selecting keypoints in the reference image. The proposed technique has been compared with local window based matching technique and found to provide better texture conservation in the resultant dense image. The proposed algorithm sets a new mechanism for computing dense depth from sparse disparity points which though requires further improvement in terms of accuracy.