Team research project · November 2025
Person-disjoint sign language recognition
Transfer learning with MobileNetV2, evaluated on people held out from training.
Conceptual illustration; metric reproduced from the report.
Research question
A gesture recogniser should be evaluated on people it has not seen during training. This project explores isolated sign recognition with an emphasis on person-disjoint evaluation.
Methodology
The team used MobileNetV2 transfer learning and five-fold cross-validation with held-out signers. The report documents the data preparation, training procedure and evaluation.
Reported results
The report records 83.3% mean accuracy and 0.812 macro-F1 under its person-disjoint five-fold protocol. These are study-specific measurements, not a claim of general sign-language understanding.
Limitations
The task concerns isolated gestures. Continuous sign language adds temporal structure, grammar and context, so these results do not establish a complete sign-language translation system.
Related research
Cognitive modelling with ACT-R