Graph Laplacian Regularization based on the Differences of Neighboring Pixels for Conditional Convolutions for Instance Segmentation

Abstract

We propose a simple and effective regularization method for instance segmentation, GLRDN-L2 (Graph Laplacian Regularization based on Differences of Neighboring Pixels). Instance segmentation is a challenging task in computer vision. For many years, ROI-based methods such as Mask R-CNN have dominantly presented the top performances; however, the recently proposed CondInst, which employs dynamic FCNs as a mask head and performs instance-aware mask prediction, outperforms Mask R-CNN. To our best knowledge, all methods optimize a model based on pixel-wise losses such as Dice Loss. Even with the results of high-resolution masks, there are problems such as blurred boundaries and hollows in the instances. We assume that these problems are due to the spatial structure and contextual information contained in the relationships between neighboring pixels not being incorporated well in the model. To address these problems, we propose a regularization that penalizes the errors in the spatial structure with a graph composed of the differences between neighboring pixels. We compare models trained with and without our regularization in CondInst to validate the effect of a natural extension that adds differentiation to the loss function and demonstrate performance improvement on both COCO and Cityscapes datasets.