Research map / 3D vision / 3D generation
SignAvatars: A Large-scale 3D Sign Language Holistic Motion Dataset and Benchmark
Zhengdi Yu, Shaoli Huang, yongkang cheng, Tolga Birdal
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add biomechanical hand constraints to the fitting objective (ablation) | SignAvatars annotation pipeline · EHF | PA-MPVPE (Hands): 5.4 -> 4.7 | Improved |
| add biomechanical hand constraints to the fitting objective (ablation) | SignAvatars annotation pipeline · EHF | MPVPE (Hands): 12.5 -> 9.7 | Improved |
| add biomechanical hand constraints to the fitting objective (ablation) | SignAvatars annotation pipeline · EHF | PA-MPJPE (Body): 16.5 -> 15.6 | Improved |
| use the optimization-based annotation pipeline instead of regression-based holistic methods | 3D holistic mesh recovery · EHF | PA-MPVPE (Holistic): 48.7 -> 12.9 (vs OSX) | Improved |
| replace pre-trained CLIP features with the parallel linguistic feature generator (PLFG) (ablation) | SignVAE · SignAvatars HamNoSys holistic subset | R-Precision top-1: 0.385 -> 0.429 | Improved |
| replace pre-trained CLIP features with PLFG (ablation) | SignVAE · SignAvatars HamNoSys holistic subset | MM-dist: 3.056 -> 2.651 | Improved |
| use VQVAE backbone instead of diffusion (MDM) with the same linguistic features (ablation) | SignVAE vs SignDiffuse · SignAvatars HamNoSys holistic subset | R-Precision top-1: 0.285 -> 0.429 | Improved |
| use 3D holistic mesh-derived motion instead of 2D cues for HamNoSys production (ablation) | Ham2Pose · SignAvatars HamNoSys holistic subset | DTW-MJE rank top-1: 0.092 -> 0.253 | Improved |
| evaluate arms only (Gesture) instead of holistic motion (ablation) | SignVAE · SignAvatars HamNoSys subset | MM-dist: FID 0.884 -> 0.851; MM-dist 2.651 -> 2.876 | Mixed |
| use word-level prompts instead of spoken-language prompts | SignVAE · SignAvatars holistic | FID: 4.359 -> 0.756 | Improved |
| train and evaluate on mesh-derived 3D keypoints instead of 3D-lifted OpenPose keypoints (ablation) | Progressive Transformer on Phoenix-2014T · Phoenix-2014T (eval model C_ours) | R-Precision top-1: 0.095 -> 0.134 | Improved |
Closest papers by meaning
- MySign: A High-Fidelity Motion-Capture Dataset for 3D Sign Generation in Bahasa Isyarat Malaysia (NeurIPS 2026 · similarity 0.92)
- Reconstructing Signing Avatars From Video Using Linguistic Priors (CVPR 2023 · similarity 0.91)
- Neural Sign Actors: A Diffusion Model for 3D Sign Language Production from Text (CVPR 2024 · similarity 0.90)
- YouTube-SL-25: A Large-Scale, Open-Domain Multilingual Sign Language Parallel Corpus (ICLR 2025 · similarity 0.88)
- Scaling Sign Language Translation (NeurIPS 2024 · similarity 0.88)
- Lost in Translation, Found in Embeddings: Sign Language Translation and Alignment (NeurIPS 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
