DINOv3 Satellite Feature Visualization

A repo to explore the DINOv3 checkpoint trained on satellite imagery using PCA visualization.

Requirements

Python Version: 3.12.11

Critical Version Notes

⚠️ transformers version is critical - The satellite DINOv3 model (facebook/dinov3-vitl16-pretrain-sat493m) requires the development version of transformers (4.56.0.dev0) for proper compatibility.

Installation

1. Install transformers development version first:

pip install git+https://github.com/huggingface/transformers.git@main

2. Install other dependencies:

pip install -r requirements.txt

Exact Dependency Versions

torch==2.7.1
torchvision==0.22.0
transformers==4.56.0.dev0
datasets==3.6.0
scikit-learn==1.7.1
matplotlib==3.10.5
numpy==2.3.2
Pillow==11.3.0

Usage

python visualize_features.py

The script will:

1. Load the satellite dataset
2. Initialize the DINOv3 satellite model
3. Extract features and perform PCA analysis
4. Generate visualizations saved to results/pca_features.png