Neural Disparity Computation for Dense Two-Frame Stereo Correspondence

This work aims at defining a new method for matching correspondences in stereoscopic image analysis. A representation of occlusions drives the overall matching process. Based on the taxonomy proposed by Scharstein and Szelinsky (2002, IJCV, 47, 7-42), the dense stereo matching process is divided into three tasks: matching cost computation, aggregation of local evidence and computation of disparity values. Within the second and third phases new strategies are introduced in an attempt to improve the reliability of results. Aggregation is based on a new local matching measure, and neural techniques compute disparities adaptively. Two experimental studies were conducted to evaluate and compare the solutions proposed. The first uses a standard well-known dataset including data with true disparity maps; the second study was conducted on complex real images acquired by a scanning electron microscope (SEM).