BallastAttN: Occlusion-Robust 3D Railway Ballast Characterization using Data Synthesis and Deep Learning

BallastAttN is a deep learning pipeline for automated 3D railway ballast characterization under the heavy occlusion typical of field conditions.

• Builds comprehensive synthetic training data from high-fidelity 3D scans of new and degraded ballast particles.
• Improves point-cloud instance segmentation through edge-aware voxelization and curriculum learning.
• Reconstructs complete particle shapes from partial point clouds for robust size and morphology measurement.
• Distinguishes new and degraded ballast using properties including 3D sphericity, Flat and Elongated Ratio, and Angularity Index.

Recommended citation: Ding, K., Fan, Q., & Tutumluer, E. (2026). "BallastAttN: Occlusion-Robust 3D Railway Ballast Characterization using Data Synthesis and Deep Learning." Transportation Research Record.
Download Paper