Semantic image segmentation using region-based object detector

Abstract

Semantic image segmentation plays a central role in image understanding and could be applied in various applications. With the development of deep learning techniques, great progress has been made in recent years. However, it is still very challenging to produce accurate segmentation results especially near object boundaries. Furthermore, all the methods with top performances are contributed by pixel-level annotations which needs expensive human effort and tremendous labeling time. In this paper, we explore a simple semantic segmentation approach using region-based object detector which only needs bounding box annotations. The main idea is using object detector to classify region proposals and then applying saliency detection method to segment such classified proposals. Experimental results on PASCAL VOC 2012 validation dataset show its comparable performance with fully supervised methods by masks.