Npcalvin/sign-bridge-model

ASL fingerspelling letters (A-Z), SigLIP2 fine-tune

SigLIP2-base (93M) fine-tuned on public ASL alphabet datasets (ASL Alphabet, Marxulia v03, ayuraj) plus webcam crops from the author. Input: 224x224 RGB crop around ONE hand, squashed, scaled to (x-0.5)/0.5. Output: 26 logits (A-Z, see config.json id2label).

Honest numbers (top-1 on a dataset NOT used for training): about 76% on a clean unseen set, 68% on average over three unseen sets. Letters J and Z involve motion and are NOT recognised by this still-image model. Trained on a few hundred images per letter per dataset; expect lower accuracy for other people, cameras and backgrounds. Check the licences of the source datasets before reuse.