International Elte Summer School in Robotics and Entrepeneurship 2025 project at the University of Southern Denmark (SDU), exploring how to detect surface scratches on LEGO bricks from a very small dataset of extreme-resolution photographs.
Team: Mahmood Seoud · Shinjae Kang · Davide Ragogna · Daniel Bösch
- ~100–200 sample images, only of scratched bricks (no clean baseline)
- 4500×4500 pixel resolution — most scratches are invisible at default zoom
- Contaminants (hair, fingerprints, dirt) easily get mistaken for scratches
- Very limited labeling budget
The approach that worked best for us was AnomalyDINO — a few-shot anomaly detection method built on DINOv2 features. With only a handful of "normal" reference crops it localizes scratches directly on the brick surface, no fine-tuning required.
| Single scratch detected | Multiple scratches detected |
|---|---|
 |
 |
| Anomaly score 0.261 — scratch localized as a clear purple streak. | Anomaly score 0.405 — several small scratches lit up near the edge. |
| Clean brick (true negative) | Failure case |
|---|---|
 |
 |
| Anomaly score 0.050 — no spurious hotspots on an unscratched brick. | Anomaly score 0.089 — the mask latches onto unrelated edge artifacts while the real scratch stays invisible. |
| Approach | Accuracy | Strengths | Weaknesses |
|---|---|---|---|
| AnomalyDINO (few-shot) | best qualitative results | High accuracy, needs only a few normal examples, no training | Still needs annotated reference crops |
| YOLOv11 (tile classification) | 74.1% / F1 0.73 | Geometry-agnostic after tiling, usable on new bricks | Heavy labeling effort per tile |
| VLM prompting (Llama-4-Maverick via Groq) | 71.7% / F1 0.53 | Zero-shot, flexible across resolutions and brick types | Low precision, inconsistent reasoning |
We first experimented with synthetic data generation in BlenderProc, then built a full CNN pipeline with custom filtering and random tile sampling, and also evaluated a vision-language model. Each had instructive failure modes that pushed us toward the few-shot approach.
The full writeup — pipeline diagrams, filtering design, tile sampling, VLM evaluation, and how to reproduce each experiment — lives in docs/APPROACHES.md.
The original presentation is also checked in: presentation.pdf.
- AnomalyDINO — the method behind our headline results
- Anomalib — broader anomaly detection toolkit
- HGNet — another interesting direction we didn't get to
- Survey on vision-based defect detection