Skip to main content

Key Results

All results are on Galaxy10 DECaLS (17,736 images, 10 classes). The primary metric is balanced accuracy (mean per-class recall), because the largest class is 7.9× the smallest. Accuracy of a classifier that always predicts the majority class would be 14.9%.

Classification​

Balanced accuracy on the test set
SimpleCNN (scratch)59.5%
DINOv2 linear probe55.5%
EfficientNet-B0 fine-tuned82.7%
DINOv2 ViT-B/14 fine-tuned83.9%
Fusion head (single run)86.3%
Provenance. Model scores in this section are copied from the output cells of the AstroVision notebooks (and, for DINOv2 per-class recall, read from an exported figure). The documentation data pack contains no metrics file for them, so they cannot be re-derived from the pack; the cell of origin is listed for each value in _facts/astrovision/FACTS_INDEX.md §3 and shown when hovering over a number.
Fine-tuning closes the domain gap

DINOv2 goes from 55.5% (frozen features) to 83.9% after 10 epochs at a learning rate of 5 × 10⁻⁶, ahead of EfficientNet-B0 (82.7%).

The fusion bar is not directly comparable

The fusion model is evaluated on a different test split (2,661 images, stratified) from the other models (2,660, random permutation), and an ablation of the same architecture shows no gain from the added modalities:

Fusion ablation (30 epochs, test)Balanced acc.Δ vs DINOv2 only
DINOv2 only85.95%+0.000
+ morphometry86.12%+0.002
+ SDSS photometry85.90%−0.001
+ morphometry + photometry85.75%−0.002
Source: data/fusion_ablation_results.json. Differences stay below half a point, comparable to the gap between two trainings of the same configuration.

Details: Late-Fusion Multimodal, Model Benchmarking.

Segmentation​

The U-Net reaches a test mIoU of 0.540 against Otsu pseudo-labels (not manual masks). Its "bulge" class is by construction the brightest 20 % of pixels of the galaxy. See U-Net Segmentation.

Uncertainty​

On the fusion model, conformal prediction at α = 0.10 gives an empirical coverage of 90.5% with a mean set size of 1.09 classes. Temperature scaling degrades calibration (ECE 0.049 → 0.120). See Uncertainty Quantification.

Hardest class​

Disturbed galaxies have the lowest recall for every model: 0.53 for fine-tuned DINOv2, 0.44 for EfficientNet-B0. See Limitations.