SML-IS: a multi-scale annotation approach to improve instance segmentation accuracy

Abstract

Instance segmentation requires precisely outlining object boundaries in images using polygonal masks. Conventional single-scale annotations often fail to represent multi-scale appearance variations, which limits segmentation accuracy and reduces the semantic richness available to learning algorithms. We introduce Small–Medium–Large Instance Segmentation (SML-IS) labelling, a multi-scale polygon annotation strategy that provides three different annotation scales for each instance within the same image. Instead of assigning a single mask, SML-IS replicates each image and applies small, medium, and large masks for every instance, thereby tripling the number of segmentation masks while keeping the original imagery unchanged. Unlike data-augmentation methods that modify raw images, SML-IS enriches only the supervision, improving label quality and scale diversity without altering the dataset size. On a custom dataset, SML-IS consistently improves mean Average Precision (mAP) by approximately 5–6% across representative models such as Mask R-CNN and YOLOv7-seg, regardless of architectural paradigm. For domain-specific verification, the proposed method was further applied to the Kvasir-SEG colorectal polyp segmentation dataset, where SML-IS demonstrated 7% mAP gain, confirming its robustness and effectiveness in medical imaging applications. Although validated in the medical domain, the method remains broadly applicable to other visual domains such as agriculture, autonomous driving, and industrial inspection, where precise boundary delineation and high-quality annotations are essential but large-scale data collection is limited.