Research map / Architectures and efficiency / RNNs and normalization
SubUNets: End-To-End Hand Shape and Continuous Sign Language Recognition
Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, 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 a second BLSTM layer (ablation) | Word SubUNet · PHOENIX-2014 test | WER (%): 43.8 -> 43.1 | Improved |
| add full frames input instead of hand patches only (ablation) | Word SubUNet · PHOENIX-2014 test | WER (%): 44.5 -> 43.1 | Improved |
| train combined SubUNets end-to-end instead of fixing pretrained weights (ablation) | Hand + Full Frame SubUNet · PHOENIX-2014 test | WER (%): 44.2 -> 42.1 | Improved |
| add Hand SubUNet hand shape CTC to Full Frame Word SubUNet (ablation) | Word SubUNet · PHOENIX-2014 test | WER (%): 43.1 -> 42.1 | Improved |
| use hand shape expert CTC instead of generic word CTC (ablation) | combined SubUNet · PHOENIX-2014 test | WER (%): 42.6 -> 42.1 | Improved |
| use hand shape expert CTC instead of generic word CTC (ablation) | combined SubUNet · PHOENIX-2014 dev | WER (%): 43.0 -> 43.1 | No clear effect |
| use Full Sum decoding instead of Viterbi (ablation) | combined SubUNet · PHOENIX-2014 test | WER (%): 42.7 -> 42.1 | Improved |
| add HMM topology with language model decoding instead of CTC topology (ablation) | combined SubUNet · PHOENIX-2014 dev | WER (%): 43.1 -> 40.8 | Improved |
| add LM topology with silence states instead of CTC topology (ablation) | combined SubUNet · PHOENIX-2014 test | WER (%): 42.1 -> 42.2 | No clear effect |
| add Hand SubUNet end-to-end alignment instead of iterative realignment | hand shape recognition · One-Million Hands | Top-1 accuracy (%): 62.8 -> 80.3 | Improved |
| use CaffeNet instead of ResNet-50 as spatial encoder (ablation) | dummy SubUNet · GPU memory capacity | max frames loadable: 35 -> 1450 | Improved |
Closest papers by meaning
- SignBERT+: Hand-Model-Aware Self-Supervised Pre-Training for Sign Language Understanding (TPAMI 2023 · similarity 0.89)
- Addressing Resource Scarcity across Sign Languages with Multilingual Pretraining and Unified-Vocabulary Datasets (NeurIPS 2022 · similarity 0.88)
- SignRep: Enhancing Self-Supervised Sign Representations (ICCV 2025 · similarity 0.88)
- SignBERT: Pre-Training of Hand-Model-Aware Representation for Sign Language Recognition (ICCV 2021 · similarity 0.88)
- Lost in Translation, Found in Embeddings: Sign Language Translation and Alignment (NeurIPS 2026 · similarity 0.88)
- KD-MSLRT: Lightweight Sign Language Recognition Model Based on Mediapipe and 3D to 1D Knowledge Distillation (AAAI 2025 · 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.
