Common visual pattern discovery and search

Abstract

Automatically discovering common visual patterns from images and videos is a useful but challenging task. On the one hand, the definition of visual patterns is rather ambiguous, it refers to the spatial composition of frequently occurring visual primitives which correspond to local features, semantic visual parts or visual objects. For example, the wheels and the body of a car could be seen as different visual primitives, while the whole car can also be seen as an individual visual primitive. On the other hand, there exhibit large variations in visual appearance and structures even within the same kind of visual pattern, which makes visual pattern discovery a very challenging task. However, since to distinguish different kinds of visual patterns from each other is a fundamental problem of many tasks in computer vision, such as pattern recognition/classification, object detection/localization, content-based image search, many studies have been introduce to solve the problem of visual pattern discovery in the literature. In this paper, we will revisit the representative studies on discovering visual patterns and discuss these methods from the view of local-feature-based and object- proposal-based visual patterns. The local-feature-based visual pattern discovery aims to mine the visual primitives that share similar spatial layout, while the semantic-patch-based visual pattern discovery aims to mine similar semantic patterns from the object proposals that are likely to contain an entire object. Then the extensive applications of visual pattern discovery are presented.