Dual-Supervised Instance Segmentation Network Combined with Priori Corner Information

Instance segmentation is a classic problem in the field of computer vision, where the goal is to classify every pixel and differentiate object instances. Mask R-CNN is one of the most advanced and efficient algorithms in the field of instance segmentation. In practice, however, Mask R-CNN algorithm does not perform well when segmenting slender parts and edge parts of the object, not only in the standard COCO dataset but also in self-made dataset. Aiming at this problem, we propose a dual-supervised network combined with priori corner information called Corner Mask R-CNN. By using corner points as a priori information to allow the network to fit the edge parts better, the details of the instance corner points can be maintained, thereby achieving more accurate results of predictions. In addition, in order to enhance the feature expression ability of the network, a Squeeze & Excitation module is added which is similar to the attention mechanism. The method proposed is more accurate than Mask R-CNN in the MS-COCO dataset. Specifically, the average precision is 1.2 percent higher than Mask R-CNN. Besides, Corner Mask R-CNN is more effective in segmenting large objects, which is 2.5 percent higher than the benchmark network.