Large scale and diverse two-dimensional flake segmentation dataset by general-purpose and labor-efficient annotation framework

Identifying 2D material flakes under a microscope is a crucial step in mechanical exfoliation. Although automated detection algorithms reduce manual search time, building generalizable foundational models still depends on large, diverse, high-quality, real-world datasets. However, manually creating such annotated datasets is laborious and time-consuming. We present the most comprehensive dataset to date for 2D flake segmentation, comprising 7454 microscopic images with approximately 30,000 annotated flake regions. It systematically captures variations in imaging conditions and flake dimensions, with size distributions spanning four orders of magnitude. To reduce labeling effort, we introduce a self-evolving annotation ecosystem that integrates active learning, semi-supervised learning, and foundation models within a batch-wise architecture. Additionally, we propose a region-aware evaluation framework that quantifies model performance via connected-region analysis, simultaneously assessing segmentation accuracy and flake count without bias. Our resources and methodology aim to improve 2D material detection and provide a practical, cost-effective framework for adaptable annotation systems.