Detection based local feature context for image classification

The use of context is shown effective for researchers. However, most of them only take context information at the visual word level without considering the context relationship of local features. For tackling this problem, a novel method is proposed by considering the detection based local feature context. Each image is represented as background and foreground. Given a position in background or foreground, to represent this position´s visual information, we use the local feature on this position as well as other local features based on angles and distances to this position. Taking use of local feature context is more discriminative and is also invariant to rotation and scale change. The local feature context can then be applied in the task of image classification. Our method is demonstrated effective by experiments on the UIUC-Sports and Caltech-101 datasets.