JLInst: Boundary-Mask Joint Learning for Instance Segmentation

Lots of methods have been proposed to improve instance segmentation performance. However, the mask produced by state-of-the-art segmentation networks is still coarse and does not completely align with the whole object instance. Moreover, we find that better object boundary information can help instance segmentation network produce more distinct and clear object masks. Therefore, we present a simple yet effective instance segmentation framework, termed JLInst (Boundary-Mask Joint Learning for Instance Segmentation). Our methods can jointly exploit object boundary and mask semantic information in the instance segmentation network, and generate more precise mask prediction. Besides, we propose the Adaptive Gaussian Weighted Binary Cross-Entropy Loss (GW loss), to focus more on uncertain examples in pixel-level classification. Experiments show that JLInst achieves improved performance (+3.0% AP) than Mask R-CNN on COCO test-dev2017 dataset, and outperforms most recent methods in the fair comparison.