Deep Guidance Decoder with Semantic Boundary Learning for Boundary-Aware Semantic Segmentation

Abstract

Image semantic segmentation is ubiquitously used in consumer electronics, such as AI Camera, which require high accuracy at the boundaries between semantic classes. To improve the semantic boundary accuracy, we propose low complexity deep-guidance decoder (DGD) networks, trained with a novel semantic boundary learning (SBL) strategy. Our ablation studies on Cityscapes and the ADE20K most-frequent 31 classes, when using different encoders and feature extractors, confirm the effectiveness of our approach. We show that the proposed DGD with SBL significantly improve the mIoU by up to 10.4% relative gain and the mean boundary F1-score (mBF) by up to 38.5%.