Computational stereo is one of the classical problems in computer vision.Numerous algorithms and solutions have been reported in recent years focusingon developing methods for computing similarity, aggregating it to obtainspatial support and finally optimizing an energy function to find the finaldisparity. In this paper, we focus on the feature extraction component ofstereo matching architecture and we show standard CNNs operation can be used toimprove the quality of the features used to find point correspondences.Furthermore, we propose a simple space aggregation that hugely simplifies thecorrelation learning problem. Our results on benchmark data are compelling andshow promising potential even without refining the solution.
展开▼
