A Noisy-Student Multi-View Ensemble with Doppler-Spatial Fusion and Angular-Margin Learning for Radar-Based Sign Language Recognition
Sign language is the primary mode of communi-cation for over 70 million deaf and hard-of-hearing individu-als worldwide. However, existing camera-based sign languagerecognition systems tend to fail in low-light conditions, raisesignificant privacy concerns in domestic and clinical settings,and degrade when the signer is occluded or improperly framed.Millimeter-wave radar offers a privacy-preserving, illumination-invariant alternative capable of capturing fine-grained micro-Doppler signatures of hand articulation, encoded as Range-Time Maps (RTMs). In this work, we introduce a radar-adaptedNoisy-Student self-training formulation together with a Spatial-Doppler six-channel input representation that fuses spatial RTMswith per-channel FFT-derived velocity maps, and we integrateangular-margin metric learning to sharpen class separation ina fine-grained label space. On the CVPR MSLR 2026 Track 2benchmark, our approach attains a Mean Class Accuracy (MCA)of 90.89%, achieving results close to the third-best entry on thecompetition leaderboard (late submission). These findings demon-strate that radar-native semi-supervised learning combined withphysics-informed feature construction can push radar-based ges-ture recognition toward the accuracy previously associated withoptical modalities, enabling practical deployment in accessibility,smart-home, and ambient-assisted-living systems
Authors
- Ritwik Raj Saxena (ORCID: https://orcid.org/0009-0001-7876-3193)
- Iwan Sandjaja
Institutions
- Petra Christian University (ID)
- Missouri University of Science and Technology (US)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-12
- DOI
- https://doi.org/10.5281/zenodo.24179153
- Primary Topic
- Hand Gesture Recognition Systems
- Type
- preprint