Context Modeling Methods for Stereo Matching and Image Parsing.

详细信息   

• 作者：Zhu ; Shengqi.
• 学历：Ph.D.
• 年：2014
• 毕业院校：The University of Wisconsin
• Department：Computer Sciences
• ISBN：9781321165418
• CBH：3635458
• Country：USA
• 语种：English
• FileSize：9957644
• Pages：159

文摘

Practical computer vision applications are often required to handle imperfect data: data acquired by non-professional devices and labeled by amateur users. These data often exhibit saturated pixels,noisy regions and texture-less areas,and are usually only partially labeled. In this thesis,we propose several context modeling methods that address imperfect input data for two fundamental computer vision problems: stereo matching and image parsing. Context modeling is the utilization of neighborhood information that may influence the way a scene or an object is perceived. It is related to how humans recognize and understand the world,and has been proved effective in many computer vision applications. In the methods proposed in this thesis,context aggregation in a large neighborhood is used to smooth image noise and to produce robust results. In addition,iterative,continuous optimization methods are used to achieve good performance,fast optimization speed and high scalability. In the stereo matching application,we propose a local linear regression approach that models the relationship between pixel intensities and disparity levels. It captures large neighborhood interactions and reduces the impact of image noise. It is shown to produce accurate disparity maps with sharp object boundaries for the Middlebury stereo dataset. It also outperforms competitive methods in real-world stereo video with regard to temporal consistency. In the image parsing application,our focus is to handle stroke labels,where a few curves are used to represent each class region. It provides less information than full labels or partial labels,and therefore is more challenging. We present two context modeling approaches with differences in how parameters are defined and how energy functions are optimized. Our synthetic experiments show the effectiveness of our methods in context propagation. Results on real-world datasets show that our method is scalable to millions of codebook bases with a comparable accuracy to other state-of-the-art methods that use fully-labeled datasets,but with a faster running speed. Our method also shows the capability for handling stroke labels and generates accurate semantic labels with consistent regions.