Research map / Large language models / Language, safety and interpretability
BSL-1K: Scaling up co-articulated sign language recognition using mouthing cues
Samuel Albanie, Gül Varol, Liliane Momeni, Triantafyllos Afouras, Joon Son Chung, Neil Fox, 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 |
|---|---|---|---|
| lower mouthing threshold to 0.5 (more, noisier training data) (ablation) | I3D Kinetics init · BSL-1K test (manually verified) | per-instance top-1 accuracy: 69.00 -> 70.61 | Improved |
| add body and hand cues to face (full-frame instead of face-crop) (ablation) | I3D · BSL-1K m.8 pose subset | per-instance top-1 accuracy: 42.23 -> 65.57 | Improved |
| add the face to body-and-hand keypoints (ablation) | Pose->Sign · BSL-1K m.8 pose subset | per-instance top-1 accuracy: 40.47 -> 49.66 | Improved |
| use I3D on raw video instead of 2D pose (ablation) | I3D vs Pose->Sign · BSL-1K m.8 pose subset | per-instance top-1 accuracy: 49.66 -> 65.57 | Improved |
| pretrain with video pose distillation instead of random init (ablation) | I3D · BSL-1K m.8 | per-instance top-1 accuracy: 39.80 -> 70.38 | Improved |
| pretrain with video pose distillation instead of Kinetics (ablation) | I3D · BSL-1K m.8 | per-instance top-1 accuracy: 69.00 -> 70.38 | Improved |
| pretrain on gesture recognition (Jester) instead of random init (ablation) | I3D · BSL-1K m.8 | per-instance top-1 accuracy: 39.80 -> 46.93 | Improved |
| use 8 sec keyword search window instead of 1 sec (ablation) | I3D Kinetics init · BSL-1K m.8 | per-instance top-1 accuracy: 60.10 -> 69.00 | Improved |
| use 8 sec keyword search window instead of 16 sec (ablation) | I3D Kinetics init · BSL-1K m.8 | per-instance top-1 accuracy: 65.91 -> 69.00 | Improved |
| use 20 frames before mouthing peak instead of 16 (ablation) | I3D Kinetics init · BSL-1K m.8 | per-instance top-1 accuracy: 59.53 -> 71.71 | Improved |
| use 20 frames before mouthing peak instead of 24 (ablation) | I3D Kinetics init · BSL-1K m.8 | per-instance top-1 accuracy: 69.00 -> 71.71 | Improved |
| mask mouth at test time for a model trained on full frames (ablation) | I3D · BSL-1K m.8 pose subset | per-instance top-1 accuracy: 65.57 -> 34.74 | Hurt |
Closest papers by meaning
- Read and Attend: Temporal Localisation in Sign Language Videos (CVPR 2021 · similarity 0.94)
- Scaling Up Sign Spotting Through Sign Language Dictionaries (IJCV 2022 · similarity 0.93)
- Lost in Translation, Found in Embeddings: Sign Language Translation and Alignment (NeurIPS 2026 · similarity 0.90)
- YouTube-SL-25: A Large-Scale, Open-Domain Multilingual Sign Language Parallel Corpus (ICLR 2025 · similarity 0.89)
- Scaling Sign Language Translation (NeurIPS 2024 · similarity 0.89)
- KD-MSLRT: Lightweight Sign Language Recognition Model Based on Mediapipe and 3D to 1D Knowledge Distillation (AAAI 2025 · similarity 0.89)
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.
