Research map / 3D vision / Human pose estimation
emg2pose: A Large and Diverse Benchmark for Surface Electromyographic Hand Pose Estimation
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 emg2pose, a large public dataset of 16-channel wrist sEMG (2 kHz) paired with motion-capture hand pose labels from 193 users, plus baselines (NeuroPose, SensingDynamics, and a new velocity-based vemg2pose) for pose regression and tracking.
Problem
Hand pose estimation from wrist sEMG lacks large, diverse public datasets, and existing sEMG pose models have been limited to few users and placements, so they generalize poorly to new users, device placements, and kinematics.
Approach
Release a large-scale sEMG-to-pose benchmark with mocap-derived joint-angle labels and three held-out generalization test sets, and propose vemg2pose, which predicts joint angular velocities with a TDS convolutional featurizer and autoregressive LSTM decoder, integrating them into angles.
Key results
- vemg2pose regression achieves 12.2 degrees angular error and 15.8 mm landmark distance on held-out users, versus 13.2 and 17.5 for NeuroPose.
- Tracking with initial pose gives 7.7 degrees and 10.3 mm on held-out users; held-out user+stage remains hardest at 11.0 degrees for tracking.
- Increasing the number of training users and stages consistently improves held-out generalization, and excluding held-out stages or users from training markedly degrades performance.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use vemg2pose instead of NeuroPose | sEMG pose regression · Held-out users | Angular Error (°): 13.2 ± 1.1 -> 12.2 ± 1.3 | Helped |
| Use vemg2pose instead of NeuroPose | sEMG pose regression · Held-out stages | Angular Error (°): 17.2 ± 1.7 -> 15.2 ± 1.6 | Helped |
| Use vemg2pose instead of NeuroPose | sEMG pose regression · Held-out users and stages | Angular Error (°): 17.5 ± 1.5 -> 15.8 ± 1.4 | Helped |
| Use vemg2pose instead of NeuroPose | sEMG pose regression · Held-out users | Landmark Distance (mm): 17.5 ± 1.3 -> 15.8 ± 1.9 | Helped |
| Use vemg2pose instead of NeuroPose | sEMG pose regression · Held-out stages | Landmark Distance (mm): 24.0 ± 2.1 -> 20.4 ± 2.2 | Helped |
| Use vemg2pose instead of NeuroPose | sEMG pose regression · Held-out users and stages | Landmark Distance (mm): 24.9 ± 1.7 -> 21.6 ± 2.0 | Helped |
| Increase the number of training users (ablation) | vemg2pose tracking · Held-out users | Error | Helped |
| Increase the number of training stages per user (ablation) | vemg2pose tracking · All modes of generalization | Error | Helped |
| Include held-out stages in training instead of excluding them (ablation) | vemg2pose tracking · Held-out stages | Error | Helped |
| Include held-out users in training instead of excluding them (ablation) | vemg2pose tracking · Held-out users | Error | Helped |
| Perform the same behaviors with visual occlusion (ablation) | sEMG pose reconstruction · Occluded versus visible hand behaviors | Accuracy | No clear effect |
| Use hand-hand or hand-object interactions (ablation) | vemg2pose tracking · Interaction versus non-interaction stages | Accuracy | No clear effect |
Surprising findings
- Visual occlusion does not affect sEMG-based pose reconstruction error, as expected for a non-vision modality.
- Predicting angular velocities rather than joint angles was sufficient for smooth predictions in the baselines.
Limitations
- Motion-capture labels suffer from occlusion (IK failed for 12.7% of frames) and wrist angles are not tracked.
- Landmark metrics depend on a default hand model that does not match each user's anatomy; models sometimes predict physically infeasible kinematics.
- Models have higher landmark errors than CV baselines, and sEMG-RD band access is limited, precluding human-in-the-loop testing.
Experimental setup
Datasets: emg2pose
Benchmarks: emg2pose
Metrics: mean absolute joint angular error (degrees) mean landmark distance (mm)
Baselines: NeuroPose SensingDynamics emg2pose (angle-prediction variant)
Builds on: NeuroPose SensingDynamics Time-Depth Separable Convolution UmeTrack
Closest papers by meaning
- emg2qwerty: A Large Dataset with Baselines for Touch Typing using Surface Electromyography (NeurIPS 2024 · similarity 0.91)
- Posture-Informed Muscular Force Learning for Robust Hand Pressure Estimation (NeurIPS 2024 · similarity 0.89)
- SpGesture: Source-Free Domain-adaptive sEMG-based Gesture Recognition with Jaccard Attentive Spiking Neural Network (NeurIPS 2024 · similarity 0.88)
- SplashNet: Split‑and‑Share Encoders for Accurate and Efficient Typing with Surface Electromyography (NeurIPS 2025 · similarity 0.87)
- MS-MANO: Enabling Hand Pose Tracking with Biomechanical Constraints (CVPR 2024 · similarity 0.87)
- EMGBench: Benchmarking Out-of-Distribution Generalization and Adaptation for Electromyography (NeurIPS 2024 · 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.