The NGT200 Dataset: Geometric Multi-View Isolated Sign Recognition
Published in GRaM Workshop @ ICML 2024, 2024
Recommended citation: Ranum, O., Wessels, D., Otterspeer, G., Bekkers, E., Roelofsen, F., Andersen, J. (2024). "The NGT200 Dataset." GRaM Workshop @ ICML 2024. https://openreview.net/forum?id=idkNzTC67X
We introduce the NGT200 dataset, a multi-view benchmark for isolated sign recognition recorded from three calibrated viewpoints by Deaf signers and a synthetic avatar. We demonstrate that viewpoint variation is a fundamental challenge for pose-based sign recognition, and propose an SE(2)-equivariant model (Temporal-PONITA) that improves accuracy by 8–22% over the state-of-the-art SL-GCN baseline.