Deep Convolutional Networks for Scene Parsing

We propose a deep learning strategy for scene parsing, i.e. to asssign a class label to each pixel of an image. We investigate the use of deep convolutional network for modeling the complex scene label structures, relying on a supervised greedy learning strategy. Compared to standard approaches based on CRFs, our strategy does not need hand-crafted features, allows modeling more complex spatial dependencies and has a lower inference cost. Experiments over the MSRC benchmark and the LabelMe dataset show the effectiveness of our approach.