Reproduce “DINO-YOLO: Self-Supervised Pre-training for Data-Efficient Object Detection in Civil Engineering Applications” on OpenResearch

About this paper

Malaisree P, Youwai S, Kitkobsin T, Janrungautai S, Amorndechaphon D, Rojanavasu P

DINO-YOLO introduces a hybrid object detection framework that integrates DINOv3 self-supervised vision transformer features into the YOLOv12 backbone, enabling data-efficient and real-time object detection for civil engineering applications. This approach yielded up to a 73.5% relative improvement in [email protected] on moderate data regimes like KITTI while maintaining real-time inference at 30-47 FPS.

Key insights

• Optimal DINO-YOLO configurations are scale-dependent, with varying DINOv3 variants and integration strategies (Single, Dual, Triple) performing best for different YOLO backbone sizes.
• Moderate data regimes (5,000-10,000 images) represent an "inflection point" where self-supervised pre-training provides transformative performance benefits (e.g., 73.5% mAP improvement).
• Extreme data scarcity (<1,000 images) and highly specialized visual domains show diminishing returns for self-supervised pre-training, indicating limits to architectural innovation alone without sufficient task-specific fine-tuning data.

Summary by alphaXiv.