Hierarchical Learning of Dominant Constellations for Object Class Recognition

The importance of spatial configuration information for object class recognition is widely recognized. Single isolated local appearance codes are often ambiguous. On the other hand, object classes are often characterized by groups of local features appearing in a specific spatial structure. Learning these structures can provide additional discriminant cues and boost recognition performance. However, the problem of learning such features automatically from raw images remains largely uninvestigated. In contrast to previous approaches which require accurate localization and segmentation of objects to learn spatial information, we propose learning by hierarchical voting to identify frequently occurring spatial relationships among local features directly from raw images. The method is resistant to common geometric perturbations in both the training and test data. We describe a novel representation developed to this end and present experimental results that validate its efficacy by demonstrating the improvement in class recognition results realized by including the additional learned information.