Abstract
ABSTRACTWe present a coarse-to-fine stereo matching optimization applicable to methods utilizing the Disparity Space Image (DSI) structure. With the Three-dimensional Recursive Search algorithm (3DRS), a coarse disparity seed is obtained first, with minimal computational effort. The coarse disparity seed is then used as a guidance to locally compute the DSI disparity space with a reduced number of disparity hypotheses, yielding significantly shorter execution times for the disparity computation. The method performance was measured on the well-known Dynamic Programming (DP) DSI-based method and the images from the Middlebury set. The DP method with the DSI optimization applied maintains or improves the overall level of disparity map accuracy while delivering a near sevenfold speed-up of execution in comparison to DP alone. We furthermore show that the optimized method's performance does not depend on the expected input disparity range, which is commonly restricted, or expected to be defined upfront, for DSI-based stereo matching methods.
Highlights
Computation of dense disparity maps is one of the most researched problems in computer vision, specially in the area of 3D modeling and reconstruction
Our work focuses on a pixel-based stereo Dynamic programming (DP) algorithm as described by Cox et al [8] and further expanded by Bobick and Intille [9] with the introduction of the Disparity Space Image (DSI) concept and Ground Control Points (GCPs)
The 3DRS, DP and the proposed hybrid 3DRS-guided dynamic programming (3GDP) algorithm have been implemented within StereoTest, a visual evaluation environment we have developed for the purpose of evaluating stereo algorithms
Summary
Computation of dense disparity maps is one of the most researched problems in computer vision, specially in the area of 3D modeling and reconstruction. A taxonomy of dense, two frame, passive stereo methods has been proposed [2] generally dividing passive methods into local, which generate disparity maps through local matching cost aggregation, or global, which aim to minimize a global energy function. Global methods, such as Graph Cuts [3] or Belief Propagation [4] generally outperform the local in terms of accuracy, but the execution time of global methods is significantly inferior, making them a poor choice for real-time applications required by recent applications such as autonomous vehicles. Some of them include using a tree-like structure which spans multiple scanlines [10], aggregating the matching cost across scanlines [15], reusing calculated paths [7] or performing a second DP pass in the vertical direction
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
