动态参数估计.pdfVIP

Multi-View Active Shape Model with Robust Parameter Estimation Li Zhang, Haizhou Ai Department of Computer Science and Technology, Tsinghua University, Beijing, 100084, China ahz@ Abstract to localize objects with large view changes, especially for the application of multi-view face alignment. These works Active Shape Model is an ef?cient way for localizing ob- often involve multiple shape and texture models built by jects with variable shapes. When ASM is extended to multi- non-linear methods such as Gaussian mixtures [1] and Ker- view cases, the parameter estimation approaches in previ- nel PCA [6], while self-occlusion can also be handled by ous works are often sensitive to the initial view, as they do learning visibility from training data [11]. However, since not handle the unreliability of local texture search, which the texture model used in the local search of each label point can be caused by bad initialization or cluttered background. depends on the view, these methods are often very sensi- To overcome this problem, we propose a novel algorithm tive to the estimation of the hidden view parameter. When for parameter estimation, using robust estimators to remove the initial view is not predicted correct, the results of lo- outliers. By weighting dynamically, our method acts as a cal search become unreliable. If the estimation of the shape model selection method, which reveals the hidden shape parameter does not deal with the potential outliers, these and view parameters from noisy observations of local tex- multi-view ASM approaches will fail. ture models. Experiments and comparisons on multi-view To overcome this problem, we propose a novel parame- face alignment are carried out to show the ef?ciency of our ter estimation approach for multi-view Active Shape Model. approa