The 3D-LEX Dataset
Published:
 | Our team recently finalized the production of 3D-LEX v1.0: 3D Lexicons for American Sign Language and Sign Language of the Netherlands, and the work is currently in preparation for publication! |
Abstract: In this work, we introduce the 3D-LEX dataset, featuring high-resolution, three-dimensional lexicons of prevalent signs in American Sign Language and Sign Language of the Netherlands. The collection includes 1000 signs from each language, encompassing handshapes, body poses, and facial features captured using three distinct motion capture techniques. Our data acquisition method demonstrates high efficiency, with the capacity to capture approximately one sign every 10 seconds. 3D-LEX supports in-depth analysis of sign features and facilitates the generation of 2D projections from any viewpoint. Additionally, the collection has been aligned with existing sign language benchmarks and resources, to support research aimed at exploring how 3D data can be leveraged to support existing video benchmarks. We showcase the utility of 3D-LEX in one envisioned benefit, namely the task of semi-automatic annotation of phonetic properties. Downstream evaluations of the new annotations in an isolated sign recognition task, using the WLASL benchmark, demonstrate that the novel labels enhance recognition accuracy by 5% compared to the absence of phonetic labels, and by 1% over annotations provided by expert linguists.
–More Information coming soon–