Research map / Label-efficient and robust learning / Recognition and long-tail
BEST: BERT Pre-training for Sign Language Recognition with Coupling Tokenization
Weichao Zhao, Hezhen Hu, Wengang Zhou, Jiaxin Shi, Houqiang Li
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 |
|---|---|---|---|
| Use coupled hand-and-body tokenization rather than separate vector quantizers (ablation) | BEST · MSASL100 | Top-1 per-instance accuracy: 77.54 -> 80.98 | Improved |
| Use coupled hand-and-body tokenization rather than K-Means (ablation) | BEST · MSASL100 | Top-1 per-instance accuracy: 74.63 -> 80.98 | Improved |
| Add hand-and-body masking instead of no pre-training (ablation) | BEST · MSASL100 | Top-1 per-instance accuracy: 73.44 -> 80.98 | Improved |
| Mask hand features only rather than body features only (ablation) | BEST · MSASL100 | Top-1 per-instance accuracy: 68.82 -> 76.75 | Improved |
| Add simultaneous hand-and-body masking rather than hand-only masking (ablation) | BEST · MSASL100 | Top-1 per-instance accuracy: 76.75 -> 80.98 | Improved |
| Increase the fraction of pre-training data (ablation) | BEST · MSASL100 | Top-1 per-instance accuracy: 74.50 -> 77.54 -> 78.47 -> 80.98 | Improved |
| Use MUM rather than the SignBERT random masking strategy (ablation) | BEST pre-training · MSASL1000 | Top-1 per-instance accuracy: 53.95 -> 58.82 | Improved |
| Use token classification rather than regression as the pre-training objective (ablation) | BEST pre-training · MSASL1000 | Top-1 per-instance accuracy: 56.90 -> 58.82 | Improved |
| Add BEST pre-training rather than train without pre-training (ablation) | BEST · WLASL2000 | Top-1 per-instance accuracy: 27.13 -> 46.25 | Improved |
| Fuse BEST pose predictions with RGB predictions | BEST · MSASL100 | Top-1 per-instance accuracy: 80.98 -> 89.56 | Improved |
| Use BEST instead of SignBERT | Pose-based sign recognition · MSASL1000 | Top-1 per-instance accuracy: 49.54 -> 58.82 | Improved |
| Use BEST instead of SignBERT | Pose-based sign recognition · WLASL2000 | Top-1 per-instance accuracy: 39.40 -> 46.25 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SignBERT — SignBERT+: Hand-Model-Aware Self-Supervised Pre-Training for Sign Language Understanding (TPAMI 2023) · better on 8 of 8 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Scaling up Multimodal Pre-Training for Sign Language Understanding (TPAMI 2025 · similarity 0.90)
- Signs as Tokens: A Retrieval-Enhanced Multilingual Sign Language Generator (ICCV 2025 · similarity 0.89)
- Uni-Sign: Toward Unified Sign Language Understanding at Scale (ICLR 2025 · similarity 0.88)
- Scaling Sign Language Translation (NeurIPS 2024 · similarity 0.87)
- Conditional Variational Autoencoder for Sign Language Translation with Cross-Modal Alignment (AAAI 2024 · similarity 0.87)
- Lost in Translation, Found in Embeddings: Sign Language Translation and Alignment (NeurIPS 2026 · similarity 0.87)
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.
