Semantic image segmentation via guidance of image classification.

This paper describes a joint segmentation and classification approach that exploits global image features to validate the predictions from local appearance descriptors and to ensure their consistent labeling. The in-between interplay is encoded by a parameter-learning process of a unified deep learning model embedding a fully convolution network portion. Although FCN has a relatively large recept field, the integration of the image content as a whole makes the prediction more reasonable and logical, since coincidences in local neighborhoods are more likely to be depressed given global structures. We also propose a content-sensitive co-occurrence priori for label compatibility, which provides additional constraints for CRF based segmentation.