recommended_min_similarity

rapidtools.processing.reid.recommended_min_similarity(model_id: str) → float

Cosine-similarity threshold that suits a backbone.

Backbones place their scores on different scales: CLIP packs both matching and non-matching vehicles into a narrow band near 1, DINOv2 spreads them out. The values are the 99th percentile of the similarity between neighbouring, different, parked vehicles in a street survey.

Parameters:

model_id – Hugging Face model identifier.

Returns:

float – Threshold at or above which two crops count as one object.

Example

>>> recommended_min_similarity('facebook/dinov2-small')
0.83
>>> recommended_min_similarity('some/other-backbone')
0.89