Co-occurrence Random Forests for Object Localization and Classification

Abstract

Learning techniques based on random forests have been lately proposed for constructing discriminant codebooks for image classification and object localization. However, such methods do not generalize well to dealing with weakly labeled data. To extend their applicability, we consider incorporating co-occurrence information among image features into learning random forests. The resulting classifiers can detect common patterns among objects of the same category, and avoid being trapped by large background patterns that may sporadically appear in the images. Our experimental results demonstrate that the proposed approach can effectively handle weakly labeled data and meanwhile derive a more discriminant codebook for image classification.