Research map / Label-efficient and robust learning / Test-time adaptation
EMGBench: Benchmarking Out-of-Distribution Generalization and Adaptation for Electromyography
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.
Introduces EMGBench, an open-source out-of-distribution benchmark for EMG gesture classification across nine datasets (including a new FlexWear-HD high-density wearable dataset), evaluating leave-one-subject-out generalization and time-series adaptation with ImageNet-pretrained CNNs and ViTs.
Problem
EMG gesture classifiers are usually evaluated with random or mixed splits that inflate accuracy, and there is no standardized benchmark for inter-subject generalization and adaptation across EMG datasets, making results hard to compare.
Approach
Standardize EMG evaluation around out-of-distribution tasks (leave-one-subject-out CV, chronological train-test splits for few-shot fine-tuning, and intersession fine-tuning) across nine public and newly collected EMG datasets, converting signals to 2D activity maps (raw heatmap, RMS, STFT, CWT) and benchmarking ImageNet-pretrained image classifiers.
Key results
- Few-shot fine-tuning on the first 20% of a held-out subject's data raises LOSO-CV accuracy substantially on most datasets (e.g., Myo Dataset from 74.5% to 95.1% with raw heatmaps).
- FlexWear-HD achieves the highest LOSO-CV accuracy among datasets, reaching 83.9% test accuracy for 10 gestures with EfficientNet before fine-tuning.
- Benchmarks show no single preprocessing or architecture wins everywhere; ResNet18, EfficientNet, ViT and EfficientViT each lead on different datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune using initial data from a left-out subject after pretraining on other subjects | EMG gesture classifiers · LOSO-CV datasets | accuracy | Helped |
| increase fine-tuning data from 5% to 20% of the left-out subject's data (ablation) | EMG gesture classifiers · Myo Dataset | accuracy: 93.2 -> 95.0 | Helped |
| increase fine-tuning data from 5% to 20% of the left-out subject's data (ablation) | EMG gesture classifiers · Ninapro DB5 | accuracy: 48.4 -> 69.7 | Helped |
| use STFT or CWT rather than raw heatmaps (ablation) | ResNet18 EMG classifier · MCS and Hyser LOSO-CV | accuracy: 67.7 -> 74.4; 44.4 -> 54.4; 67.7 -> 77.9; 44.4 -> 58.0 | Helped |
| replace raw heatmaps with RMS heatmaps (ablation) | ResNet18 EMG classifier · Myo Dataset LOSO-CV | accuracy: 74.5 -> 68.7 | Hurt |
| replace raw heatmaps with STFT or CWT representations (ablation) | ResNet18 EMG classifier · FlexWear-HD LOSO-CV | accuracy: 77.7 -> 75.3; 77.7 -> 75.1 | Hurt |
| fine-tune on the first session of the left-out subject (ablation) | EMG gesture classifiers · UCI EMG intersession | accuracy: 79.8 -> 93.5 | Helped |
| fine-tune on the first session of the left-out subject (ablation) | EMG gesture classifiers · CapgMyo intersession | accuracy: 47.9 -> 62.1 | Helped |
| fine-tune on the first session of the left-out subject (ablation) | EMG gesture classifiers · Hyser intersession | accuracy: 72.5 -> 60.1 | Hurt |
| fine-tune on the first session of the left-out subject (ablation) | EMG gesture classifiers · FlexWear-HD intersession | accuracy: 81.7 -> 99.1 | Helped |
| use phase-based spectrogram or HHT phase representation instead of magnitude-based preprocessing (ablation) | Ninapro DB5 classifier · LOSO-CV | accuracy | Hurt |
| use IRM or CORAL instead of standard cross-entropy training (ablation) | EMG gesture classifiers · generalization and few-shot fine-tuning | accuracy | No clear effect |
Surprising findings
- Phase-based representations (STFT phase, HHT phase) perform far worse than magnitude-based representations on Ninapro DB5 (about 22-31% vs roughly 70% accuracy).
- Domain generalization methods IRM and CORAL perform comparably to standard cross-entropy training.
Limitations
- Evaluated activity maps only up to 224x224 resolution.
- Datasets are cue-based controlled gestures and do not capture real-world spontaneous gestures.
- Several datasets lack transition data, and LOSO-CV accuracy remains low on some datasets such as Ninapro DB5 and DB2.
Experimental setup
Datasets: Myo Dataset UCI EMG Ninapro DB5 CapgMyo DB-b Ninapro DB2 Ninapro DB3 Multi-channel sEMG (MCS) Hyser FlexWear-HD
Benchmarks: EMGBench
Metrics: accuracy AUROC
Baselines: ResNet18 EfficientNet ViT EfficientViT invariant risk minimization correlation alignment
Builds on: ImageNet pretraining ResNet Vision Transformer EfficientNet invariant risk minimization correlation alignment
Closest papers by meaning
- emg2qwerty: A Large Dataset with Baselines for Touch Typing using Surface Electromyography (NeurIPS 2024 · similarity 0.88)
- emg2pose: A Large and Diverse Benchmark for Surface Electromyographic Hand Pose Estimation (NeurIPS 2024 · similarity 0.87)
- SplashNet: Split‑and‑Share Encoders for Accurate and Efficient Typing with Surface Electromyography (NeurIPS 2025 · similarity 0.85)
- SpGesture: Source-Free Domain-adaptive sEMG-based Gesture Recognition with Jaccard Attentive Spiking Neural Network (NeurIPS 2024 · similarity 0.84)
- Quasi-Linear ICA for Motor Unit Decomposition during Dynamic Contractions (NeurIPS 2026 · similarity 0.84)
- Revisiting Noise Resilience Strategies in Gesture Recognition: Short-Term Enhancement in sEMG Analysis (ICML 2025 · similarity 0.82)
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.