Ranking Prior Likelihood Distributions for Bayesian Shape Localization Framework

• Authors: Shuicheng Yan;Mingjing Li;Hongjiang Zhang;Qiansheng Cheng
• Venue: ICCV '03 Proceedings of the Ninth IEEE International Conference on Computer Vision - Volume 2
• Year: 2003

In this paper, we formulate the shape localization problem in theBayesian framework. In the learning stage, we propose theConstrained Rank Boost approach to model the likelihood of localfeatures associated with the keypoints of an object, like face,while preserve the prior ranking order between the ground truthposition of a keypoint and its neighbors; in the inferring stage, asimple efficient iterative algorithm is proposed to uncover the MAPshape by locally modeling the likelihood distribution around eachkey point via our proposed variational Locally Weighted Learning(VLWL) method. Our proposed framework has the following benefits:1) compared to the classical PCA models, the likelihood presentedby the ranking prior likelihood model has more discriminating poweras to the optimal position and its neighbors, especially in theproblem with ambiguity between the optimal positions and theirneighbors; 2) the VLWL method guarantees that the posteriorprobability of the derived shape increases monotonously; and 3) theabove two methods are both based on accurate probabilityformulation, which spontaneously leads to a robust confidencemeasure for the discovered shape. Moreover, we present atheoretical analysis for the convergence of the ConstrainedRank-Boost. Extensive experiments compared with the Active ShapeModels demonstrate the accuracy, robustness, and stability of ourproposed framework.