HamNoSys-Guided Benchmark for Fine-Grained Handshape Recognition in Sign Language

A new standard for detailed handshape recognition in sign language, based on the Hamburg Notation System (HamNoSys), has been established. This dataset includes 144,000 RGB images sourced from 15 individuals, encompassing 160 handshape categories as outlined in the official HamNoSys 4 Handshapes Chart. The research examines appearance-based models (ResNet-18 and ViT-B/16) alongside landmark-based models (graph convolutional network and XGBoost) using two methods: a class-stratified subject-dependent split and a 15-fold leave-one-subject-out (LOSO) approach. Additionally, the same model types are tested on external datasets (LSWH100 and ASL Fingerspelling Dataset A) for further context. This study fulfills the demand for extensive, phonetically defined visual inventories that incorporate signer-aware evaluations in computational sign-language tasks. The full paper can be found on arXiv (2608.10588).