Apollo clapperboard classifier โ€” 2026.10.16.7

Scores eBay listings for Apollo: the probability that a listing is a production-used original film/TV clapperboard (not a replica, novelty item, popcorn bucket, blank, or photo of a slate).

File What it is
siglip2_so400m_256_vision.onnx SigLIP 2 SO400M/16 @256 vision tower, fp32 ONNX, pixel_values [N,3,256,256] โ†’ L2-normalised embedding [N,1152]
apollo_head.npz Trained head (numpy): per-photo logistic regression, title character n-gram TF-IDF, and a combiner over photo scores, title score and price. meta holds the version and alert thresholds.

Preprocessing: letterbox to a white square, bilinear resize to 256ร—256, scale to [-1, 1] (src/machine/classifier.py:prepare in the Apollo repo).

Results (grouped 5-fold cross-validation on 6,637 hand-labelled eBay listings, 287 production-used)

PR-AUC ROC-AUC Precision @95% recall Precision @99% recall Alerts to catch 99%
0.992 1.000 0.95 0.90 319

Alert thresholds shipped with this version: REVIEW โ‰ฅ 0.21 (highest threshold that alerted on every real slate), FOUND โ‰ฅ 0.95 (best F1). Full report: reports/README.md in the Apollo repo.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for maymeridian/apollo-clapperboard

Quantized
(4)
this model