Research map / Label-efficient and robust learning / Self-supervised and contrastive
SignRep: Enhancing Self-Supervised Sign Representations
Ryan Wong, Necati Cihan Camgoz, Richard Bowden
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 sign-prior pretraining instead of pixel MAE on YT-SL (ablation) | Hiera-B MAE · WLASL2000 recognition | instance top-1: 57.2 -> 59.9 | Improved |
| pretrain on YT-SL with pixel MAE instead of Kinetics MAE (ablation) | Hiera-B · WLASL2000 recognition | instance top-1: 51.5 -> 57.2 | Improved |
| add class probability distribution loss (kappa 0.2) (ablation) | SignRep · WLASL2000 recognition | instance top-1: 59.9 -> 61.0 | Improved |
| raise class distribution loss weight kappa from 0.2 to 0.5 (ablation) | SignRep · WLASL2000 recognition | instance top-1: 61.0 -> 59.8 | Hurt |
| add masking during pretraining (ablation) | SignRep · WLASL retrieval | DCG: 46.3 -> 48.5 | Improved |
| add variance-covariance regularization (ablation) | SignRep · WLASL retrieval | DCG: 48.5 -> 49.9 | Improved |
| add adversarial style loss (ablation) | SignRep · WLASL retrieval | DCG: 49.9 -> 50.7 | Improved |
| combine angle, keypoint and distance priors rather than keypoints alone (ablation) | SignRep · WLASL retrieval | DCG: 32.9 -> 48.5 | Improved |
| combine all three priors rather than distance prior alone (ablation) | SignRep · WLASL retrieval | DCG: 47.2 -> 48.5 | Improved |
| use hand-activity weighted averaging of segment features instead of plain average (ablation) | SignRep retrieval · ASL-Citizen retrieval | Rec@1: 37.47 -> 49.95 | Improved |
| use sign-prior pretrained features instead of hand joint angles for retrieval | SignRep · ASL-Citizen retrieval | Rec@1: 8.81 -> 49.95 | Improved |
| replace DinoV2 LoRA backbone with frozen SignRep features | Sign2GPT · CSL-Daily | BLEU-4: 12.96 -> 16.33 | Improved |
Closest papers by meaning
- Addressing Resource Scarcity across Sign Languages with Multilingual Pretraining and Unified-Vocabulary Datasets (NeurIPS 2022 · similarity 0.92)
- Learning Effective Sign Features without Text for Gloss-free Sign Language Translation (CVPR 2026 · similarity 0.91)
- SignBERT+: Hand-Model-Aware Self-Supervised Pre-Training for Sign Language Understanding (TPAMI 2023 · similarity 0.91)
- Improving Gloss-free Sign Language Translation by Reducing Representation Density (NeurIPS 2024 · similarity 0.89)
- Lost in Translation, Found in Embeddings: Sign Language Translation and Alignment (NeurIPS 2026 · similarity 0.89)
- Uni-Sign: Toward Unified Sign Language Understanding at Scale (ICLR 2025 · 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.
