Image-Level Supervised Instance Segmentation Using Instance-Wise Boundary

Abstract

Recently, most image-level supervised instance segmentation methods extend Class Attention Maps (CAMs) to find the entire instance masks. Inter-pixel Relation Network (IRNet) can effectively generate the class-wise boundary maps for attention score propagation. However, class-wise boundary is likely to cause the failure of segmentation among instances. In this work, we find instance-wise information can be extracted from the displacement field of IRNet. Motivated by the observations, an improved IRNet-based instance segmentation method with instance-wise boundary has been developed. Experimental results based on PASCAL VOC 2012 demonstrate the effectiveness of our proposed method. Compared with the recent state-of-the-art methods, the mean average precision can be increased by 4.3% without any additional annotations.