Abstract
In this paper, we propose a novel semantic segmentation of panoramic images. It classifies panoramic images at the pixel level, to assign a semantic tag to each pixel of the image to recognize different objects or regions in the image. We combine the characteristics of semantic segmentation algorithm to improve HoHoNet algorithm. To solve the problem of large size of the panoramic image, HarDNet is used for the backbone, which improves the segmentation accuracy as well as the training efficiency. The ECA attention mechanism is embedded to address the loss caused by the varying importance of different channels in the feature map during the convolutional pooling process. Finally, the Boundary loss function is used to make the proposed method pay more attention to the edge features of the object and further improve the effect and accuracy of panoramic semantic segmentation. Finally, the improved method was evaluated on the Stanford 2D-3D-Semantics dataset, and its mAcc and mIoU reached 54.1% and 66.7%, respectively. The experimental results show that the segmentation ability and real-time performance of our algorithm are superior to the current panoramic semantic segmentation network.