HDNet: Hybrid Distance Network for Semantic Segmentation

Abstract Semantic segmentation is currently solved as a pixel-wise labeling task, which predicts the label of each pixel based on its features. However, current methods isolate the relations of points in a feature map and cause the discontinuous segmentation results. In order to solve this problem, we propose a hybrid distance network to measure the distance from two aspects. First, the Hybrid Distance Relation is proposed to model the relations between a point and its context regions to capture contexts in a feature map by an elegant combination of positional distance and high-dimension feature distance. Then, a Location Aware Attention module is proposed to efficiently sample the contexts by the positional distance and produces sparse Hybrid Distance Relations. It synthesizes the different contexts of each point and generates position-wise attention value to compact object-level representation. During the training step, High-dimension Feature Distance loss is also presented as an auxiliary loss to compact category-level representation in feature space. Experiments show that the proposed HDNet achieves state-of-the-art performance with interpretability and efficiency on three challenging semantic segmentation benchmarks: Pascal Context, ADE20K, and COCO Stuff 10K.