Localizing Occluders with Compositional Convolutional Networks

Compositional convolutional networks are generative models of neural network features, that achieve state of the art results when classifying partially occluded objects, even when they have not been exposed to occluded objects during training. While previous results showed the potential of CompositionalNets at localizing occluders, this remains to be confirmed quantitatively. In this work, we study the performance of CompositionalNets at localizing occluders in an image. We propose to extend the original model with a mixture of von-Mises-Fisher distributions. We show that this extension increases the model's ability to localize occluders in an image while retaining an exceptional performance at classifying partially occluded objects.