Multi-plane Image Stitching Base on Image Semantic under Large Viewpoint Changes

Abstract

Traditional image stitching method is under the assumption that the image content is approximately in a same plane, and it will result in severe stitching distortion for images with large viewpoint changes. To solve the problem above, a robust multi-plane image stitching method based on image semantic is proposed in this paper. Assuming that the regions with the same semantic label are in the same plane, the image can be divided into multiple planes with transition borderline by semantic segmentation. Finally, image stitching is completed by mixing multiple image planes with gaussian weights along the borderline, and warping image with perspective and similarity transformation. The experimental results indicate that the proposed method performs more accurately and naturally, and is robust to large viewpoint changes.