Semantic-level Fusion with Noise in Computer Vision.

详细信息   

• 作者：Xu ; Jiang.
• 学历：Doctor
• 年：2011
• 导师：Wu, Ying,eadvisorKatsaggelos, Aggelos K.ecommittee memberPappas, Thrasyvoulos N.ecommittee member
• 毕业院校：Northwestern University
• Department：Electrical and Computer Engineering
• ISBN：9781124861005
• CBH：3469776
• Country：USA
• 语种：English
• FileSize：4224251
• Pages：114

文摘

Semantic-level fusion is a critical and common problem in many computer vision tasks. When some of the estimates are wrong, the fusion problem remain unsolved. Two scenarios of fusing erroneous estimates are investigated in this paper: One is the fusion when the estimates have multiple modes, and the other is the fusion when only pairwise relations can be extracted between estimates. In the first scenario, a large number of multimodal partial estimates exist, and the fusion generally becomes complicated and intractable, thus it is desirable to find an effective and scalable fusion method to integrate these partial estimates. I present a novel and effective approach to fusing multimodal partial estimates in a principled way. In this new formulation, fusion of partial estimates is related to a computational geometry problem of finding the minimum-volume orthotope. An effective branch-and-bound search algorithm is designed to find the global optimal solution. The method is tested on both numerical cases and real cases tracking articulated and occluded objects), and compared to RANSAC method. The proposed algorithm outperforms other approaches. In the second scenario, when only pairwise relations can be extracted between estimates, the fusion problem can be converted to a grouping problem. However, existing grouping methods cannot deal with many erroneous estimates. Most affinity-based grouping methods only model the inclusive relation among the data. When the data set contains a significant amount of noisy data that should not be included in any clusters, these methods are likely to lead to undesired results. To address this issue, I present a new approach called bipolar grouping that is targeted on extracting the groups from the data while excluding the noise. This new approach incorporates both inclusive and exclusive relations among data. I also propose applications including discovering common objects in images, tracking targets in clutter and motion segmentation. In general, semantic-level fusion with noise is an important but complicated problem in computer vision. In my work, I investigate two common scenarios of fusing erroneous estimates. I have found noise-robust algorithms for both scenarios, and the method can be applied in many applications.