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

The research introduces Small–Medium–Large Instance Segmentation (SML-IS), a multi-scale annotation strategy for improving instance segmentation accuracy. Traditional single-scale annotations often fail to capture multi-scale appearance variations, limiting segmentation performance and reducing semantic richness for learning algorithms.

SML-IS overcomes this limitation by providing three distinct annotation scales – small, medium, and large – for each instance within an image. Rather than assigning a single mask, SML-IS creates multiple copies of the image and applies the small, medium, and large masks to every instance, effectively tripling the number of segmentation masks while preserving the original image integrity.

Distinguishing from data augmentation techniques that modify raw images, SML-IS enhances only the supervisory information. This annotation enrichment approach improves label quality and scale diversity without increasing the dataset size. On a custom dataset, the method consistently raised the mean Average Precision (mAP) by around 5–6% across popular models like Mask R-CNN and YOLOv7-seg, irrespective of architectural design.

To validate its performance in medical imaging, SML-IS was applied to the Kvasir-SEG colorectal polyp segmentation dataset, where it achieved a 7% mAP improvement. The robustness and effectiveness of the method were confirmed in this application.

While SML-IS has been validated in the medical domain, the research emphasizes its broad applicability to various visual domains. Applications range from agriculture and autonomous driving to industrial inspection, where precise boundary delineation and high-quality annotations are crucial, but large-scale data collection can be challenging.

The study was financially supported by the Recep Tayyip Erdogan University Development Foundation (Grant number: 02026002024140) and acknowledges their crucial contribution to the research. The authors are affiliated with the Departments of Electrical and Electronics Engineering at Ataturk University (Turkey), the Recep Tayyip Erdogan University (Turkey), and the University of Cambridge (UK).

The research is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution, and reproduction with proper attribution to the original authors and the source.