Online Hard Region Mining for Semantic Segmentation

摘要

Recent advances in semantic segmentation have made significant progress by enlarging the reception fields or capturing contextual information. Semantic segmentation is considered as a per-pixel classification problem. Hard discriminate region existing in an image will limit segmentation accuracy. In this work, we propose an approach to increase the attention to local semantic segmentation performance by region-based hard region mining. To analyse the performance on three popular semantic segmentation datasets, including PASCAL VOC 2012, PASCAL Context and Camvid, we experiment two different semantic segmentation networks, Deeplab v3 and FCN. Our experimental results show consistent improvement, which demonstrating the efficacy of our approach.