NGT200: Geometric Multi-View Isolated Sign Recognition
Published:
| We introduce NGT200, a multi-view isolated sign recognition dataset for Sign Language of the Netherlands (NGT), designed to support research at the intersection of geometric deep learning and sign language processing. The dataset features synchronized multi-camera recordings with 3D pose annotations, enabling viewpoint-invariant recognition approaches. Paper: openreview.net/forum?id=idkNzTC67X Venue: ICML 2024 Workshop on Geometry-grounded Representation Learning and Generative Modeling Authors: Ranum, O., Wessels, D., Otterspeer, G., Bekkers, E., Roelofsen, F., Andersen, J. |
📄 Paper · 🌐 Project Page · 🗄️ Dataset (OSF) · 💻 Code
Abstract: We present the NGT200 dataset, featuring geometric multi-view recordings of 200 signs from Sign Language of the Netherlands. The dataset is designed to benchmark viewpoint-invariant and geometry-aware recognition methods, bridging the gap between 3D geometric deep learning and real-world sign language video recognition. Multi-view synchronized recordings are paired with 3D skeletal pose annotations, providing a resource for evaluating isolated sign recognition under varying viewpoints and spatial conditions.