Finding Pictures of Objects in Large Collections of Images

摘要: Retrieving images from very large collections, using image content as a key, is becoming an important problem. Users prefer to ask for pictures notions of that are strongly oriented the presence abstractly defined objects. Computer programs implement these queries automatically desirable, but hard build because conventional object recognition techniques computer vision cannot recognize general objects in contexts. This paper describes our approach recognition, which structured around sequence increasingly specialized grouping activities assemble coherent regions can be shown satisfy stringent constraints. The constraints satisfied provide form classification quite view distinguished by: far richer involvement early visual primitives, including color and texture; hierarchical learning strategies process; ability deal with rather uncontrolled configurations We illustrate properties four case-studies: one demonstrating use texture descriptors; showing how trees described by fusing geometric properties; scenery concepts grouped features; this yields program tell, accurately, whether picture contains naked people or not.