Cue Integration by Similarity Rank List Coding - Application to Invariant Object Recognition

Similarity rank lists provide a method for learning generalization of classifiers from examples. Here, we apply it to invariant object recognition and demonstrate that it performs better than other approaches on view and illumination invariant recognition. Recognition from a single view reaches 87% success rate. To study its real world capabilities we introduce subsqare rank matching that works on image patches and RUBJECTS100, a database of 100 objects under varying pose and illumination, and a set of natural scenes containing these objects.