Research map / Architectures and efficiency / Attention and transformers
Revisiting Noise Resilience Strategies in Gesture Recognition: Short-Term Enhancement in sEMG Analysis
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes STEM, a sliding-window attention module with weight sharing that captures short-term sEMG features, and STET, a transformer combining it with long-term attention and self-supervised sensor-wise masking pretraining, for noise-robust gesture recognition and joint-angle regression.
Problem
Gesture recognition from surface EMG is degraded by real-world noise (skin condition changes, interference, electrode issues), and existing sEMG models treat signals as generic time series, neglecting short-term features and noise robustness.
Approach
Add a learnable short-term enhancement module (STEM) that uses sliding-window self-attention with weight sharing in parallel with a long-term attention decoder, fused for classification, plus a sensor-wise, geometric-segment masked reconstruction pretraining and asymmetric loss.
Key results
- STET achieves 90.76% overall accuracy on GRABMyo gesture classification, versus 87.71% for Informer and 80.43% for Zerveas et al.
- On Ninapro DB2 joint-angle regression, STET reaches PCC 0.877 and NRMSE 0.073 versus 0.869 and 0.076 for sBERT-OHME.
- Under additive Gaussian noise, Transformer's accuracy drop falls from 25% to 10% when STEM is added; the paper claims noise impact on STET is reduced by more than 20% versus best competitors.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add EIPC pretraining (ablation) | Transformer gesture classifier · GRABMyo overall | accuracy (%): 85.73% to 86.33% (+0.60%) | Helped |
| Add the long-term decoder instead of the fully connected decoder (ablation) | Transformer gesture classifier · GRABMyo overall | accuracy (%): +1.16% | Helped |
| Add the short-term decoder instead of the fully connected decoder (ablation) | Transformer gesture classifier · GRABMyo overall | accuracy (%): +1.39% | Helped |
| Fuse long-term and short-term features (ablation) | STET gesture classifier · GRABMyo overall | accuracy (%): +2.46% | Helped |
| Use Asymmetric Loss instead of Cross Entropy Loss (ablation) | STET gesture classifier · GRABMyo overall | accuracy (%): 0.73% | Helped |
| Add STEM/STET design to the Transformer | Transformer · Additive Gaussian noise | accuracy drop rate (%): 25% -> 10% | Helped |
| Add STEM/STET design to the Transformer | Transformer · Multiplicative Gaussian noise | accuracy drop rate (%): 16% -> 10% | Helped |
| Add STEM/STET design to Informer | Informer · Signal loss | accuracy drop rate (%): 26% -> 17% | Helped |
| Add STEM to Informer | Informer · GRABMyo overall | accuracy (%): 87.71% to 89.14% | Helped |
| Use STET instead of Zerveas et al. (2021) | Transformer-based gesture recognition · GRABMyo overall | accuracy (%): 80.43% to 90.76% | Helped |
| Add STEM with weight sharing | Transformer · Transformer inference | inference time (GPU): 3.8 ms -> 3.9 ms | No clear effect |
| Remove weight sharing from STEM (ablation) | Transformer with STEM · Transformer inference | parameter count: 489233 -> 581137 | Hurt |
Surprising findings
- Adding STEM with weight sharing increases parameters only ~1.7% (481,169 to 489,233), whereas non-shared attention raises them ~21%.
- Informer+STEM improves Rest-class accuracy from 83.56% to 90.33%, while the STET-only Rest accuracy is 95.33%.
Limitations
- Experiments are participant-dependent (user-specific), so models are not shown to generalize across users without per-user training.
- Noise robustness is evaluated with synthetic noise injected into lab data; real-world evaluation is limited to a small human-subject study with eight participants.
- Pretraining robustness ablation was conducted on only three subjects due to time constraints.
Experimental setup
Datasets: GRABMyo Ninapro DB2
Benchmarks: GRABMyo Ninapro DB2
Metrics: accuracy standard deviation Pearson correlation coefficient NRMSE average curvature epoch time accuracy drop rate under noise
Baselines: Asif et al. TCN GRU TEMGNet Zerveas et al. Informer LST-EMG-Net LSTM BERT sBERT-OHME
Builds on: Transformer Informer Asymmetric Loss Focal Transformer
Closest papers by meaning
- SpGesture: Source-Free Domain-adaptive sEMG-based Gesture Recognition with Jaccard Attentive Spiking Neural Network (NeurIPS 2024 · similarity 0.88)
- emg2qwerty: A Large Dataset with Baselines for Touch Typing using Surface Electromyography (NeurIPS 2024 · similarity 0.87)
- emg2pose: A Large and Diverse Benchmark for Surface Electromyographic Hand Pose Estimation (NeurIPS 2024 · similarity 0.86)
- SplashNet: Split‑and‑Share Encoders for Accurate and Efficient Typing with Surface Electromyography (NeurIPS 2025 · similarity 0.84)
- Posture-Informed Muscular Force Learning for Robust Hand Pressure Estimation (NeurIPS 2024 · similarity 0.83)
- GeneOH Diffusion: Towards Generalizable Hand-Object Interaction Denoising via Denoising Diffusion (ICLR 2024 · similarity 0.83)
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.