Curvature-based pebble segmentation for reconstructed surface meshes

Abstract. Accurate segmentation of pebbles in complex 3D scenes is essential to understand sediment transport and river dynamics. In this study, we present a curvature-based instance segmentation approach for detecting and characterizing pebbles from 3D surface reconstructions. Our method is validated using high-resolution ground truth models, allowing for a quantitative assessment of segmentation accuracy. The workflow involves reconstructing a sandbox scene using available open-source or commercial software packages, segmenting individual pebbles based on curvature features, and evaluating segmentation performance using detection metrics, primary axes estimation, 3D orientation retrieval, and surface area comparisons. Results show a high detection precision (0.980), with segmentation errors primarily attributed to under-segmentation caused by overly smooth surface reconstructions. Primary axis estimation via bounding box fitting proves more reliable than ellipsoid fitting, particularly for the A and B axes, while the C-axis remains the most challenging due to partial occlusion. 3D orientation estimation reveals variability, with cumulative errors ranging from less than 5° to more than 45°, highlighting the difficulty in retrieving full orientations from incomplete segments. Surface area metrics indicate that our approach prioritizes precision over recall, with nine out of ten ground truth pebbles achieving IoU values above 0.8. In addition, we introduce a Python-based segmentation tool that provides detailed morphological and color-based metrics for each detected pebble. Our findings emphasize the advantages of true 3D analysis over traditional 2D photo-sieving approaches and suggest future improvements through refined segmentation algorithms and enhanced surface reconstructions.