BAProto: Boundary-Aware Prototype for High-quality Instance Segmentation.

Yuxuan Zhang,Wei Yang,Rong Hu
DOI: https://doi.org/10.1109/icme55011.2023.00398
2023-01-01
Abstract:To date, the boundary quality remains unsatisfactory in instance segmentation, resulting in increasing attention to the mask refinement mechanism. Such a mechanism is supposed to be accurate, efficient and generic to the existing models. Yet, few methods met the three factors simultaneously. In this paper, to address this issue, we propose a Boundary-Aware Prototype (BAProto) for boundary refinement, which conducts pixel-wise prediction through the similarity of boundary representation and the specific prototype. Such a prototype is obtained by a memory unit for comprehensive learning. To our best knowledge, BAProto is the first approach that satisfies the above three factors at the same time. In particular, we elaborately design a three-phase segmentation loss, focusing on the learning of different regions to extract discriminating boundary representations for prototype establishment. Extensive experimental results show that BAProto is precise, efficient and model-agnostic on COCO and Cityscapes datasets.
What problem does this paper attempt to address?