Automatic Visual Class Formation using Image Fragment Matching

Low–level vision approaches, such as local image features, are an important component of bottom–up machine vision solutions. They are able to effectively identify local visual similarities between fragments of underlying physical objects. Such vision approaches are used to build a learning system capable to form meaningful visual objects out of unlabelled collections of images. By capturing similar fragments of images, the underlying physical objects are extracted and their visual appearances are generalized. This leads to formation of visual objects, which (typically) represent specific underlying physical objects in a form of automatically extracted multiple template images.