SAMSnake: A generic contour-based instance segmentation network assisted by Efficient Segment Anything Model.

Abstract

Contour-based instance segmentation has gained significant attention due to its efficiency and ability to produce precise segmentation boundaries. In this paper, we propose SAMSnake, a novel contour-based instance segmentation network. Our method introduces three key innovations: (1) We modify the structure of existing contour segmentation methods, decouple the detector from the traditional contour segmentation framework, and enhance the flexibility of downstream tasks. (2) We improved the existing contour initialization method by introducing the EfficientSAM and building the EfficientSAM-based Contour Initialization (ECI) module, which generates an initial contour closely aligned with the true instance boundary. (3) We integrate Dynamic Matching Loss (DML) and normalization offsets into the iterative optimization process, forming the Normalization Contour Optimization (NCO) module, which refines contour deformation with high precision. Additionally, we incorporate heatmap and boundary map supervision to enhance training stability further. Extensive experiments on benchmark datasets, including Cityscapes, Semantic Boundaries Dataset (SBD), COCO, KITTI INStance (KINS), and COCOA, demonstrate that SAMSnake achieves state-of-the-art performance. Our approach achieves 36.4% mAP on Cityscapes, 61.4% mAP on SBD, 38.8% mAP on COCO, 36.7% mAP on KINS, and 46.0% mAP on COCOA. Our implementation is available at https://github.com/Giansar-Wu/SAMSnake.