Many state-of-the-art approaches for object recognition reduce the problem to a 0-1 classifi cation task. Such re- ductions allow one to leverage sophisticated classifi ers for learning. These models are typically trained independentl