A Large-Scale 3D Representation Dataset and Benchmark for Continuous Sign Language Understanding

American Sign Language (ASL) is a visually rich 3D language essential to the Deaf and Hard-of-Hearing (D/HH) communities, where 3D mesh representations have increasingly surpassed traditional skeletal poses in capturing expressive detail and articulatory nuance. Yet, the scarcity of large-scale 3D ASL datasets continues to constrain model generalization and the enhancement of cross-linguistic knowledge. To address this gap, we present a large-scale, expressive, and privacy-conscious 3D mesh dataset comprising 250+ hours of data curated from various sign language datasets¹. Building upon this resource, we establish comprehensive benchmarks for ASL translation and generation, facilitating standardized evaluation across tasks. In addition, we propose a full-body fine-grained refinement method, FusePose, which jointly integrates high-quality hand and body representations to improve articulation fidelity. Extensive experiments demonstrate that even a small amount of high-quality data can foster better cross-linguistic generation. Furthermore, we show a downstream application in controllable character animation, highlighting the broader impact of our dataset and methodology on 3D ASL research. Our experiments further show that fused hand-body representations provide stronger performance for both Sign Language Translation and Sign Language Generation/Production.