What is AstroVision?
AstroVision is an open-source pipeline for galaxy morphology on Galaxy10 DECaLS (17,736 RGB images of 256×256 pixels, 10 classes). It is the image-domain companion of AstroSpectro, which works on stellar spectra. Beyond classification, it asks whether features learned without any astronomy data also encode measurable physical properties of galaxies.
Do self-supervised visual representations (DINOv2 ViT-B/14) encode physical galaxy structure — concentration, colour, stellar population — or only appearance? The evidence collected so far is correlational: see Physical Interpretation and Limitations.
Key results
_facts/astrovision/FACTS_INDEX.md §3 and shown when hovering over a number.The per-model comparison, with its test-set caveats, is on Model Benchmarking. The multimodal late-fusion model reaches 86.3% in a single run, but its ablation shows that the extra modalities add no measurable gain over the DINOv2 branch alone — see Late-Fusion Multimodal.
What the pipeline contains
| Stage | What it does | Page |
|---|---|---|
| Data | Galaxy10 DECaLS (HDF5), seed-42 split, SDSS DR16 cross-match | Data |
| Image classifiers | SimpleCNN, EfficientNet-B0, DINOv2 linear probe and fine-tuning | Models |
| Fusion | DINOv2 features + morphometry (CAS, Gini, M20) + SDSS colours | Late-Fusion |
| Segmentation | Otsu pseudo-labels → U-Net (background / disk / bright core) | U-Net |
| Morphometry | Non-parametric CAS / Gini / M20 on Otsu and on attention masks | Morphometry |
| Uncertainty | MC Dropout, temperature scaling, conformal prediction | Uncertainty |
| Interpretation | Grad-CAM, DINOv2 attention maps, colour and population links | Attention |
Scores come from the project notebooks' output cells; the documentation data pack contains no metrics file for them. Each value carries its origin (hover over a number, or read _facts/astrovision/FACTS_INDEX.md). Models were not all evaluated on the same test images, and some headline numbers from earlier descriptions of the project do not survive a comparison with the ablation: the discrepancies are listed in the data registry rather than smoothed over.