Edge-enhanced instance segmentation by grid regions of interest

This paper focuses on the instance segmentation task. The purpose of instance segmentation is to jointly detect, classify and segment individual instances in images, so it is used to solve a large number of industrial tasks such as novel coronavirus diagnosis and autonomous driving. However, it is not easy for instance models to achieve good results in terms of both efficiency of prediction classes and segmentation results of instance edges. We propose a single-stage instance segmentation model EEMask (edge-enhanced mask), which generates grid ROIs (regions of interest) instead of proposal boxes. EEMask divides the image uniformly according to the grid and then calculates the relevance between the grids based on the distance and grayscale values. Finally, EEMask uses the grid relevance to generate grid ROIs and grid classes. In addition, we design an edge-enhanced layer, which enhances the model’s ability to perceive instance edges by increasing the number of channels with higher contrast at the instance edges. There is not any additional convolutional layer overhead, so the whole process is efficient. We evaluate EEMask on a public benchmark. On average, EEMask is 17.8% faster than BlendMask with the same training schedule. EEMask achieves a mask AP score of 39.9 on the MS COCO dataset, which outperforms Mask RCNN by 7.5% and BlendMask by 3.9%.