Benchmark: cross-signer generalization goes 22% to 59% as training signers go 1 to 4

#2
by CLERC-IO - opened
CLERC org

One signer does not generalize. Four start to.

We ran a small BiLSTM on EPEE, our open ASL benchmark (600 clips, 4 native Deaf signers, 150 parallel phrases, sign-level gloss annotations): tested on a signer held entirely out of training, accuracy goes 22% -> 40% -> 50% -> 59% as you train on 1 -> 2 -> 3 -> 4 signers. Macro-F1 0.13 -> 0.38. And the curve is not saturated.

Most sign language corpora are single-signer, interpreter-based, or unstructured video. Inter-signer variation is the problem to model.

Dataset (CC BY-NC-SA, DOI 10.5281/zenodo.20268565): https://huggingface.co/datasets/CLERC-DATA/epee
Full method and honest caveats in BENCHMARK.md.

Sign up or log in to comment