Semantic segmentation using high order information
Parvin Razzaghi · 2017
In this paper, a new approach to semantic segmentation is proposed. Semantic segmentation provides a semantic label for each pixel in image using a predefined set of labels. Most pixel labeling approaches use Conditional Random Field (CRF). The main goal of this paper is to incorporate high level information in pixel labeling. In our approach, high level information as well as low level information is incorporated in data term of CRF. In the proposed approach, the set of semantic labels is divided into three subsets based on effective high level information for each class label. Then, a unified approach based on CRF is designed for pixel labeling for different categories of class labels. In our approach, the prior knowledge of segment labels, segment-level, object-level and context-level information are jointly considered. Prior probability of segment labels is considered by employing foreground/background information to lead each segment to be in a true class label category. Object detection results and shape models are used to consider object-level information. Here, object detection results and shape information are combined in a way shape model compensates object detector's false negatives. Also, to consider context-level information, context model is introduced. Co-occurrences of semantic labels, their spatial layout with respect to each other and spatial extent of each class label are jointly considered in context model. Finally, to optimize the problem, graph cut is used. To evaluate the proposed approach, it is applied on the well-known 21-MSRC dataset. The obtained results show that our approach effectively incorporates the high level information in baseline CRF.