Research map / Video / Action recognition
Re-Sign: Re-Aligned End-To-End Sequence Modelling With Deep Recurrent CNN-HMMs
Oscar Koller, Sepehr Zargaran, Hermann Ney
Check 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 |
|---|---|---|---|
| Iteratively re-align training frame labels using the CNN-HMM | CNN on PHOENIX 2014 · PHOENIX 2014 Dev | WER: 33.7 -> 29.0 | Improved |
| Iteratively re-align training frame labels using the CNN-HMM | CNN-2BLSTM on PHOENIX 2014 · PHOENIX 2014 Test | WER: 32.9 -> 26.8 | Improved |
| Use full-frame images instead of tracked right-hand patches (ablation) | GoogLeNet recognition model · PHOENIX 2014 Dev | WER: 38.3 -> 33.7 | Improved |
| Add one LSTM layer to the CNN (ablation) | GoogLeNet on PHOENIX 2014 · PHOENIX 2014 Dev | WER: 33.7 -> 44.2 | Hurt |
| Initialize the CNN-LSTM with a CNN trained on PHOENIX 2014 (ablation) | CNN-LSTM on PHOENIX 2014 · PHOENIX 2014 Dev | WER: 44.2 -> 33.8 | Improved |
| Use two bidirectional LSTM layers instead of one (ablation) | CNN-BLSTM on PHOENIX 2014 · PHOENIX 2014 Dev | WER: 30.2 -> 27.1 | Improved |
| Increase each BLSTM layer from 512 to 1024 hidden units (ablation) | CNN-2BLSTM on PHOENIX 2014 · PHOENIX 2014 Dev | WER: 27.9 -> 27.1 | Improved |
| Use two BLSTM layers with 1024 units per layer rather than the one-layer CNN-LSTM setup (ablation) | End-to-end CNN recurrent model · PHOENIX 2014 Dev | WER: 29.5 -> 27.1 | Improved |
| Apply iterative re-alignment | CNN-2BLSTM on SIGNUM · SIGNUM Test | WER: 6.5 -> 4.8 | Improved |
| Increase the number of re-alignment iterations from two to three (ablation) | Full-frame CNN on PHOENIX 2014 · PHOENIX 2014 Dev | WER: 30.7 -> 29.0 | Improved |
| Increase the number of re-alignment iterations from three to four (ablation) | CNN-2BLSTM on PHOENIX 2014 · PHOENIX 2014 Dev | WER: 27.1 -> 27.2 | No clear effect |
Closest papers by meaning
- Recurrent Convolutional Neural Networks for Continuous Sign Language Recognition by Staged Optimization (CVPR 2017 · similarity 0.90)
- Weakly Supervised Learning with Multi-Stream CNN-LSTM-HMMs to Discover Sequential Parallelism in Sign Language Videos (TPAMI 2019 · similarity 0.90)
- Deep Sign: Enabling Robust Statistical Continuous Sign Language Recognition via Hybrid CNN-HMMs (IJCV 2018 · similarity 0.89)
- Iterative Alignment Network for Continuous Sign Language Recognition (CVPR 2019 · similarity 0.88)
- Deep Hand: How to Train a CNN on 1 Million Hand Images When Your Data Is Continuous and Weakly Labelled (CVPR 2016 · similarity 0.88)
- Lost in Translation, Found in Embeddings: Sign Language Translation and Alignment (NeurIPS 2026 · similarity 0.86)
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.
