Polar Edge Distance Loss in Edge-aware Plug-and-play Scheme for Semantic Segmentation

Semantic segmentation is a classic and fundamental computer vision problem dedicated to assigning each pixel with its corresponding class. Some recent methods introduce edge-based information for improving the segmentation performance. However these methods are specific and limited to certain network architectures, and they cannot be transferred to other models or tasks. Therefore, we propose a universal edge supervision method, EPS (Edge-aware Plug-and-play Scheme), which can be easily integrated into any semantic segmentation model. EPS extracts Edge of Ground Truth (Edge GT) with a predefined thickness from the training data and copies the decoder head as an auxiliary task for edge supervision. We introduce a novel Polar Edge Distance (PED) loss, which has lower computational complexity, is easier to optimize, and improves model performance more than other boundary-based losses. We train various recent open-source semantic segmentation models, such as Segmentation Transformer (SegFormer) and Image Cascade Network (ICNet), on the relatively simple public dataset Cityscapes and the challenging public dataset ADE20K to validate the effectiveness of the EPS and PED loss. Specifically, they can further enhance the mIoU (mean Intersection over Union) of segmentation based on SOTA (state-of-the-art) models. Moreover, we find that our method shows more significant improvements when applied to single-part closed objects with larger pixel values. EPS and PED loss can be applied to semantic and instance segmentation, with potential for future exploration in edge segmentation and weakly-supervised semantic segmentation.