Research map / Large language models / Language, safety and interpretability
Scaling Up Sign Spotting Through Sign Language Dictionaries
Gül Varol, Liliane Momeni, Samuel Albanie, Triantafyllos Afouras, Andrew Zisserman
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 |
|---|---|---|---|
| replace classification loss with contrastive Watch-Read-Lookup MIL-NCE on the MLP (ablation) | I3D BSL-1K+BslDict + MLP · BSL-1K sign spotting (Train 1064, Seen 264) | mAP: 37.13 -> 47.93 | Improved |
| add subtitles (Read) on top of Watch-Lookup (ablation) | Watch-Read-Lookup · BSL-1K sign spotting (Train 800, Unseen 264) | mAP: 11.02 -> 14.86 | Improved |
| use multiple-instance bags instead of single-instance InfoNCE (ablation) | Watch-Lookup · BSL-1K sign spotting (Train 1064, Seen 264) | mAP: 43.59 -> 44.72 | Improved |
| extend dictionary vocabulary from 800 training words to full 9k (ablation) | Watch-Read-Lookup · BSL-1K sign spotting (Train 800, Unseen 264) | mAP: 14.86 -> 15.82 | Improved |
| train on dictionary only instead of joint BSL-1K and BslDict (ablation) | I3D classification · BSL-1K sign spotting (Train 1064, Seen 264) | mAP: 19.81 -> 2.68 | Hurt |
| lower mouthing confidence threshold from 0.9 to 0.5 (more training data) (ablation) | Watch-Lookup · BSL-1K sign spotting (Train 1064) | mAP: 37.55 -> 43.65 | Improved |
| reduce test-time sliding window stride from 16 to 4 frames (ablation) | Watch-Lookup · BSL-1K sign spotting (Train 1064) | mAP: 31.96 -> 44.92 | Improved |
| stride 1 versus stride 2 (ablation) | Watch-Lookup · BSL-1K sign spotting (Train 1064) | mAP: 45.39 -> 43.65 | No clear effect |
| remove class-balanced batch sampling (ablation) | Watch-Lookup · BSL-1K sign spotting (Train 1064) | mAP: 43.65 -> 41.65 | No clear effect |
| add domain-specific MLP layers (ablation) | Watch-Lookup · BSL-1K sign spotting (Train 1064) | mAP: 43.65 -> 43.58 | No clear effect |
| discard sign synonyms from negatives (ablation) | Watch-Lookup · BSL-1K sign spotting (Train 1064) | mAP: 43.65 -> 45.03 | No clear effect |
| replace I3D trunk with S3D (ablation) | sign recognition on BSL-1K · BSL-1K recognition | per-instance top-1: 75.51 -> 64.76 | Hurt |
Closest papers by meaning
- BSL-1K: Scaling up co-articulated sign language recognition using mouthing cues (ECCV 2020 · similarity 0.93)
- Read and Attend: Temporal Localisation in Sign Language Videos (CVPR 2021 · similarity 0.91)
- Addressing Resource Scarcity across Sign Languages with Multilingual Pretraining and Unified-Vocabulary Datasets (NeurIPS 2022 · similarity 0.90)
- Automatic Dense Annotation of Large-Vocabulary Sign Language Videos (ECCV 2022 · similarity 0.90)
- Lost in Translation, Found in Embeddings: Sign Language Translation and Alignment (NeurIPS 2026 · similarity 0.88)
- SignGraph: A Sign Sequence is Worth Graphs of Nodes (CVPR 2024 · 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.
