In this paper, we investigate a novel reconfigurable part-based model, namelyAnd-Or graph model, to recognize object shapes in images. Our proposed modelconsists of four layers: leaf-nodes at the bottom are local classifiers fordetecting contour fragments; or-nodes above the leaf-nodes function as theswitches to activate their child leaf-nodes, making the model reconfigurableduring inference; and-nodes in a higher layer capture holistic shapedeformations; one root-node on the top, which is also an or-node, activates oneof its child and-nodes to deal with large global variations (e.g. differentposes and views). We propose a novel structural optimization algorithm todiscriminatively train the And-Or model from weakly annotated data. Thisalgorithm iteratively determines the model structures (e.g. the nodes and theirlayouts) along with the parameter learning. On several challenging datasets,our model demonstrates the effectiveness to perform robust shape-based objectdetection against background clutter and outperforms the other state-of-the-artapproaches. We also release a new shape database with annotations, whichincludes more than 1500 challenging shape instances, for recognition anddetection.
展开▼
