Generic object classifiers based on real image selection from the web

In this paper we present our semi-supervised technique for building object category classifiers using real image data from the Internet. Our technique not only reduces the overhead of manual training by humans, but also achieves robust classifiers that can be evaluated in real time. Given a sample object and its name (keyword), we collect a large amount of object-related images from two main image sources: Google Images and the LabelMe website. We deal with the problem of separating good training samples from noisy images by performing two steps: similar image selection and non-real image filtering. We use a variant of Gaussian Discriminant Analysis (GDA) to filter out non-real images (drawings, cartoons, etc.) that tentatively affect classifier performance in real environments. In order to select true object-related training samples, we introduce a Simile Selector Classifier (SSC) that is constructed from a small set of images taken from the sample object. The SSC not only is able to select similar samples from the large unordered set of images, but also it can separate desired object category images from other categories that have the same name (polysemes), i.e. "apple" as a fruit, or as a company logo. Finally, the experiments which we performed in real environments demonstrate the performance of our object classifiers.