haugland-master-2026

Code accompanying the master's thesis Uncertainty-Aware Crack Detection with Vision Foundation Models (Vebjørn Haugland, NTNU, 2026).

What is and isn't here

The DNV crack dataset is private property of DNV and is not redistributed in this repo. To run the pipeline end-to-end you additionally need:

• the DNV crack image+mask dataset (private)
• the DINOv3 model code: https://github.com/facebookresearch/dinov3
• the SAT-493M checkpoint dinov3_vitl16_pretrain_sat493m-eadcf0ff.pth from the DINOv3 release
• ~45 GB free disk for the precomputed per-patch feature grids that data/build_train_grids.py writes

Repo layout

haugland-master-2026/
  data/        DINO feature extraction 
  analysis/    standalone DINO-only thesis figures (chapter 4)
  models/      training, evaluation, plotting (chapters 5–6)
  figures/     outputs of analysis/ scripts + auto-generated tables
  results/     outputs of training + eval (weights ignored; JSON+PNG+logs tracked)