Exploiting deep convolutional network and patch-level CRFs for indoor semantic segmentation

Recent advances in semantic segmentation have mostly been achieved by utilizing deep convolutional neural networks (CNNs). In this paper, a novel indoor semantic segmentation method is proposed by integrating CNN and patch-level Conditional random fields (CRF). Multi-scale images are sent to CNN to capture objects of different sizes as well as to extract features at multiple scales. Patch-level CRF is constructed to further refine the object boundaries localization accuracy. Extensive experiments on the publicly available NYU V2 database demonstrate that the proposed method could obtain state-of-the-art accuracy in terms of four evaluation metrics.