CV
Selected Projects
 | SignGPT A UKRI EPSRC Programme Grant (University of Surrey, Oxford, UCL) building generative transformers for bidirectional BSL–English translation, combining computer vision, sign linguistics, and machine learning to model the full structure of signed communication. Project Page | SignGPT Project |
| Seeing Space Spatial grammar toolkit | Seeing Space — Directional Verbs and Topographic Placement Computational toolkit for spatial and productive constructions in sign languages: directional verb detection with grammatical-person labelling and a learned pose-based placement module for pronominal indexing. Active Development | Project Page (Coming soon) |
 | What's the Point? — Spatial Grammar & Index Resolution for Sign Language Processing Modular framework for spatial indexing in BSL: a pose-based index detector combined with an online entity memory that substantially improves pointing-sign recovery in CSLR without retraining the backbone. Under review | Project Page (Coming soon) |
 | NGT200 — Geometric Multi-View Isolated Sign Recognition Multi-view benchmark for pose-based isolated sign recognition of Sign Language of the Netherlands (NGT). SE(2)-equivariant Temporal-PONITA improves recognition by 8–22% over SL-GCN baseline. Paper (ICML 2024 Workshop) | Project Page | Dataset |
 | 3D-LEX v1.0 — 3D Lexicons for ASL and NGT High-resolution motion capture dataset of 2,000 signs across ASL and Sign Language of the Netherlands, with semi-automatic phonetic annotation that matches or exceeds expert accuracy. Paper (LREC-COLING 2024) | Project Page | Dataset |
 | SE(3)-Equivariant Neural Rendering: Github Repository Equivariant neural rendering is a complex task that builds upon geometric deep learning, computer vision and representation theory. One critical component of this task is scene representation ... |
 | Indoor scene generation using Point-E and Training Free Layout Control: Github Repository We present framework for extending point cloud diffusion models to accomodate indoor scene generations directly from text prompts. |
Media
 |