Objection Classification Based on Estimation of Foreground Distribution

This paper presents a novel methodology for recognizing object categories based on the estimation of image foreground distribution. We offer insight into how the matting technique can be applied in the image retrieval field and how spatial cues can be incorporated into bag-of-features approach in a proper way. The spatial cues indicate distribution of foreground objects or background environments. In our approach, a formalized framework is offered: an image descriptor is defined as a weighted concatenation of features extracted on different patches of an image, and the weights here are set to be foreground distribution values in corresponding patches. Experiments on challenging object categorization image sets demonstrate the effectiveness of this framework.