Mud Spot Ballast Segmentation

This project focused on improving ballast particle segmentation under muddy field conditions.

• Designed realistic synthetic materials for muddy ballast scenes.
• Trained a DETR-based detector and fine-tuned SAM on the augmented dataset.
• Improved segmentation performance by about 5% in both mAP and mAR.
• Validated the workflow on field data with 4.18% MAE in Fouling Index and 3.45% MAE in Particle Size Distribution.