Affinity Propagation for Class Exemplar Mining

This paper focusses on the problem of locating object class exemplars from a large corpus of images using affinity propagation. We use attributed relational graphs to represent groups of local invariant features together with their spatial arrangement. Rather than mining exemplars from the entire graph corpus, we prefer to cluster object specific exemplars. Firstly, we obtain an object specific cluster of graphs using a similarity propagation based graph clustering (SPGC) method. Here a SOM neural net based tree clustering method is used to incrementally cluster a large corpus of local invariant descriptors. The popular affinity propagation based clustering algorithm is then individually applied to each object specific cluster. Using this clustering method, we obtain object specific exemplars together with a high precision for the data associated with each exemplar. The strategy adopted is one of divide and conquer, and this greatly increases the efficiency of mining exemplars. Using the exemplars, we perform recognition using a majority voting strategy that is weighted by nearest neighbor similarity. Experiments are performed on over 80K images spanning ~500 objects, and demonstrate the performance in terms of efficiency, scalability and recognition.