Answers are grounded in this paper and cite their sources.
Mexican Sign Language (LSM) lacks large-scale corpora, and its recognition studies rarely state whether test signers appear in training. On a single LSM corpus, we measure the gap between random clip partitioning and signer-independent evaluation. Signs are represented by anatomically normalized MediaPipe keypoints and classified with a Transformer encoder. On 1664 clips of 22 signs from five identified signers, we evaluated seven architectures and nine ablations under leave-one-signer-out rotation, with ten seeds per configuration. The Transformer reached 96.1% under random partitioning but 31.5–81.9% on held-out signers, a mean gap of 32.9 percentage points (pp; 95% confidence interval 30.4–35.5). A corpus audit linked the failure on the least favorable signer to an inconsistent assignment of the dominant hand to MediaPipe’s handedness channels; a label-free canonicalization raised mean signer-independent accuracy by 5.6 pp (gap: 25.7 pp). Random partitioning compressed the differences between architectures. Permuting frame order did not reduce accuracy, and, with the number of training clips fixed, training on three of the corpus signers instead of one improved accuracy by 16.9 pp. In this corpus, the diversity of signers and of their recording conditions limits generalization more than the choice of architecture.
No takes yet. Share an insight, caveat, or question.
Becerra-Aguayo et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: