An Improved IRNet for Instance Segmentation Based on Image-level Supervision

To reduce the dependence on high-cost pixel-level annotations, instance segmentation based on image-level annotations have received researchers' attention. Inter-pixel Relation Network (IRNet) is one of the representative methods which obtain the final instance segmentation results by exploiting inter-pixel relations in Class Activation Maps (CAMs). In this paper, we introduce a contrastive loss function to improve the displacement field prediction results and a Siamese network is developed to obtain scale-invariant class boundary maps. With the improved displacement field and class boundary maps, the final segmentation results have been improved. Compared with the baseline IRNet, the mean average precision has increased by 3.2% on the PASCAL VOC 2012 val set.