John Winn, Antonio Criminisi, and Thomas Minka
This paper presents a new algorithm for the automatic recognition of object classes from images (categorization). Compact and yet discriminative appearance-based object class models are automatically learned from a set of training images. The method is simple and extremely fast, making it suitable for many applications such as semantic image retrieval, web search, and interactive image and video editing. It classifies a region according to the proportions of different visual words (clusters in feature space). The specific visual words and the typical proportions in each object are learned from a segmented training set. Each visual word is described by a mixture of Gaussians in feature space, and is constructed by greedily merging an initially large dictionary. We present a novel statistical measure of discrimination which is optimized by each merge. High classification accuracy is demonstrated for nine object classes on photographs of real objects viewed under general lighting conditions, poses and viewpoints. The set of test images used for validation comprise: i) photographs acquired by us, ii) images from the web and iii) images from the recently released Pascal dataset. Interestingly, our algorithm performs well on both texture-rich objects (e.g. grass, sky, trees) and structure-rich ones (e.g. cars, bikes, planes).
In Proc. IEEE Intl. Conf. on Computer Vision (ICCV).