Reproduce “DINOv2 Rocks Geological Image Analysis: Classification, Segmentation, and Interpretability” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Florent Brondolo, Samuel Beaussant

This study evaluates DINOv2, a self-supervised vision transformer, for classification, segmentation, and interpretability in analyzing micro-CT images of geological rock samples. The research demonstrates DINOv2's high accuracy in rock type classification and superior multi-class segmentation performance, particularly with limited training data, surpassing traditional and supervised deep learning methods.

Key insights

  • A frozen DINOv2 model inherently captures meaningful petro-morphological features, enabling accurate clustering and classification of distinct rock types in its latent space, even for out-of-distribution geological data.
  • Parameter-efficient fine-tuning (LoRA) of the DINOv2 backbone significantly boosts segmentation performance, outperforming more complex segmentation heads on frozen features and demonstrating superior data efficiency.
  • DINOv2, particularly when fine-tuned, exhibits strong generalization and robustness in handling class imbalance and limited training data, making it well-suited for data-scarce geoscientific applications.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.