Compute Morphometry

One image

from src.astrovision.morphometry import GalaxyMorphometry

m = GalaxyMorphometry(image_uint8, bg_pct=10.0, smooth_sigma=3.0)   # (H, W, 3) or (H, W)
stats = m.compute_all()
# {'C', 'A', 'S', 'Gini', 'M20', 'r_petrosian', 'sersic_n', 'n_pixels', 'cx', 'cy'}

Statistics are NaN when the Otsu mask has fewer than 10 pixels (20 for the Petrosian radius and the Sérsic fit).

A batch

from src.astrovision.morphometry import batch_morphometry

df = batch_morphometry(images, labels, max_n=5000)      # random sub-sample, seed 42
print(df.groupby('class_name')[['C', 'Gini', 'M20']].median())

batch_morphometry samples at most max_n images (default 5,000 — the reference table has 5,000 rows) and records g10_idx, label and class_name; failing images give a row of NaN. Expect roughly 10–20 minutes for 5,000 images on a desktop CPU (the notebooks report this order of magnitude).

Plots

from src.astrovision.morphometry import plot_cas_diagram, plot_morphometry_by_class
plot_cas_diagram(df, save_path='figures/seg_cas_diagram.png')
plot_morphometry_by_class(df)

Attention-guided variant

from src.astrovision.attention_morphometry import AttentionMorphometry
am = AttentionMorphometry(dino_model, device='cuda', percentile=60.0)    # fine-tuned DINOv2 ViT-B/14
df_attn = am.batch_compute(images, labels)

This uses the DINOv2 attention map as the galaxy mask. Its C is defined differently from GalaxyMorphometry.concentration; do not mix the two tables (Morphometry, Attention).

Troubleshooting

Symptom	Cause
Many NaN rows	the Otsu mask is empty or tiny (faint or very small target)
Sérsic n near 1 for all classes	known property of the simplified radial-profile fit
Slow	the loop is pure Python per image; subsample with `max_n`