Inference and Learning with Hierarchical Shape Models

摘要: In this work we introduce a hierarchical representation for object detection. We represent an in terms of parts composed contours corresponding to boundaries and symmetry axes; these are turn related edge ridge features that extracted from the image. We propose coarse-to-fine algorithm efficient detection which exploits nature model. This provides tractable framework combine bottom-up top-down computation. learn our models training images where only bounding box is provided. automate decomposition category into contours, discriminatively cost function drives matching image using Multiple Instance Learning. Using shape-based information, obtain state-of-the-art localization results on UIUC ETHZ datasets.