HTC-BI: Hybrid Task Cascade with Boundary Information for Instance Segmentation
Thi Ngoc Thuy Le, Thanh Duc Ngo · 2023
Hybrid Task Cascade (HTC) is an effective cascading architecture for instance segmentation. By fully utilizing the interdependent relationship between detection and segmentation, it has shown remarkable performance. However, exploring the potential of boundary information remains open. In this work, we propose a novel method to integrate multi-level boundary information from global to local context into the HTC framework, which entirely supports segmentation tasks at all levels and produces more accurate predictions. Our method consists of two major components: (eq1) at the image level, learning the boundary representation and combining it with the semantic segmentation branch for better spatial context information; (eq2) at the instance level, adopting boundary-preserving mask head replacing the basic mask head that jointly learns object mask and boundary. By additionally learning the boundary representation, our method achieves 39.2% AP and outperforms the original HTC on MS COCO dataset. Code and models are available at https://github.com/thuyltncs/HTCBI.