Models & Methods
AstroVision benchmarks 5 modelling strategies in increasing order of complexity:
| Model | Balanced accuracy | Key feature |
|---|---|---|
| SimpleCNN (baseline) | 59.5% | 3 conv layers from scratch |
| EfficientNet-B0 | 82.7% | Transfer learning from ImageNet (full fine-tuning) |
| DINOv2 (linear probe) | 55.5% | Frozen ViT-B/14 features, zero-shot |
| DINOv2 (fine-tuned) | 83.9% | Full ViT-B/14 fine-tuning — best image-only model |
| Late-fusion multimodal | 86.3% | DINOv2 + morphometrics + SDSS (different split, see below) |
Balanced accuracy on each model's test set. The first four rows share one split (2,660 test images); the late-fusion model uses another stratified split (2,661 test images), so the step from 83.9% to 86.3% compares different test sets. In the 30-epoch ablation the four feature configurations stay within 0.4 points of one another (86.0% for DINOv2 alone, 85.7% with morphometrics and photometry): the gain comes from the MLP head on frozen features, not from the added modalities. The fine-tuned backbone may also have seen part of the fusion test images, a leak that cannot be checked from the exported facts.
The late-fusion model combines:
- DINOv2 visual features (768-dim from ViT-B/14)
- Non-parametric morphometrics (CAS, Gini, M20)
- SDSS photometric colours (u-g, g-r, r-i, i-z)
Navigate to each model's page for implementation details and analysis.