A Two-Pipeline Instance Segmentation Network via Boundary Enhancement for Scene Understanding

For applications such as autonomous vehicles, instance segmentation of captured images about surrounding environments is critical, and in the meantime, these types of tasks require high accuracy as well. However, most existing models do not consider the extra computation caused by complex architectures and the poor ability of overly lightweight structures to capture semantic information. To enable real-time instance segmentation including semantic segmentation and object detection, a new framework based on an encoder–decoder architecture is proposed, which differs from existing work in two aspects: 1) it adopts two branches to generate a prototype and semantic masks and predicts instance class confidence, bounding box and the coefficients of prototype masks and 2) it adds a spatial attention module (SAM) and the Laplacian operator into a loss function for the segmentation branch to refine boundaries. The instance masks are then produced by combining the generated segmentation mask and mask coefficients. An expansion crop training strategy is employed to accelerate training and detection speed and alleviate bad results due to convolutional translation invariance. Extensive experiments are carried out on available public datasets, including Cityscapes Urban and COCO 2017. Results show that the proposed model achieves 38.9 AP on the Cityscapes test set and 46.8 AP on the COCO test set with eight categories of instances. Compared with other state-of-the-art methods for real-time applications, our method reaches 29.0 FPS with improved accuracy.