부산대학교 도서관

This paper proposes a direction convolution (DC) with direction information added to solve the semantic segmentation problem. In the existing semantic segmentation, the edge of the detected object is not clear. To solve this problem, the model was trained by adding edge information. Also, coord convolution is used to add pixel location information, and DC is proposed to detect the edge accurately. DC is a method to add information about eight directions around a pixel for the model to train affinity for neighboring pixels. Using the cityscape dataset, this paper’s experimental results show the usefulness of the proposed method.