Over-segment-based Self-adaptive Stereo Matching

Stereo vision is one of the most active research areas in computer vision. It is made up ofcalibration, matching and reconstruction. Meanwhile, as the most challenging topic in stereovision, matching is the fundament of other vision problems. It plays key role in the developmentof computer vision.This dissertation mainly focuses on stereo matching problem. In summary, the maincontributions are listed as followed:This dissertation reviews the basic concepts of markov random fields (MRF) and discussestwo optimization methods: belief propagation and graph cut. Besides, it revisits the high orderMRF model and its optimal methods including the field of experts and robust pnpotts model. Wepropose a self-adaptive state space reduction method to overcome the low efficiency of field ofexperts. For robust pnpotts model, a novel approaches is developed for stereo matching.Experimental results show that high order MRF can catch more interaction between pixels.This dissertation presents a novel object-oriented stereo matching on multi-scale superpixelsto generate a low-resolution depth map. It can preserve the original image structure information.This dissertation introduces a novel over-segment based self-adaptive stereo matchingalgorithm which works with MRF modeling and energy minimization framework. Colorsegmentation information is used to handle untextured region and occlusion area. Each segmentis modeled as a plane. A plane fit technique is used to make the algorithm robust to outliers. Theoptimal disparity and refined data term is approximated by applying iterative belief propagation.Experimental results on the Middlebury stereo benchmark show that this algorithm isstate-of-the-art.