CPEP: Contrastive Pose-EMG Pre-training Enhances Gesture Generalization on EMG Signals

Conference: NeurIPS 2025 arXiv: 2509.04699 Code: None Area: Human Understanding / Gesture Recognition Keywords: EMG signals, EMG, gesture recognition, contrastive learning, zero-shot classification, cross-modal alignment

TL;DR

This paper proposes the CPEP framework, which employs contrastive learning to align low-quality EMG signal representations with high-quality hand pose representations, endowing the EMG encoder with pose-awareness. CPEP is the first to achieve zero-shot recognition of unseen gestures from EMG signals, yielding a 21% improvement on in-distribution gesture classification and a 72% improvement on unseen gesture classification.

Background & Motivation

Background: Vision-based gesture recognition has matured considerably, yet remains constrained by power consumption and privacy concerns in wearable device deployments. Surface EMG (sEMG) signals are low-power and easy to integrate, making them well-suited for real-time gesture prediction on wearable platforms.

Limitations of Prior Work: (a) EMG signals exhibit low signal-to-noise ratios and high variability, limiting the effectiveness of conventional self-supervised pre-training; (b) supervised methods (e.g., emg2pose for pose regression) generalize poorly and cannot recognize unseen gestures or adapt to new users; (c) large-scale EMG data collection is costly and difficult.

Key Challenge: EMG is a "weak modality" from which high-quality representations are difficult to learn in isolation, whereas hand pose is a "strong modality" that encodes rich structural and semantic information. The core challenge is how to leverage the strong modality's prior knowledge to improve representations of the weak modality.

Goal: To enable the EMG encoder to learn pose-aware representations, supporting zero-shot gesture classification via embedding-space retrieval against pose references.

Key Insight: Inspired by CLIP-style cross-modal contrastive pre-training, with design adaptations specific to the EMG–pose setting—pre-training unimodal encoders to reduce paired data requirements, and freezing the strong modality encoder while training only the weak modality encoder.

Core Idea: Contrastive learning is used to pull EMG representations toward paired pose representations, enabling zero-shot gesture recognition without task-specific fine-tuning.

Method

Overall Architecture

CPEP consists of three stages: (1) MAE-based self-supervised pre-training of both EMG and pose encoders; (2) contrastive pre-training with the pose encoder frozen, aligning the [CLS] representations of both modalities via InfoNCE; (3) downstream evaluation via linear probing or zero-shot nearest-neighbor retrieval.

Key Designs

1. Unimodal Encoder Pre-training (MAE):

   • Function: Pre-trains separate Transformer encoders for EMG and pose modalities independently.
   • Mechanism: Standard MAE with temporal patching, mask ratio \(r=50\%\); only unmasked tokens are encoded, and the decoder reconstructs the full sequence. \(\mathcal{L}_{\text{MAE}} = \frac{1}{|\mathcal{M}|}\sum_{i\in\mathcal{M}} \|\psi(\phi(\{\mathbf{z}_j\}_{j\notin\mathcal{M}}))_i - \mathbf{z}_i\|_2^2\)
   • Design Motivation: Learning robust unimodal representations prior to contrastive alignment reduces the paired data requirements during the contrastive stage.

2. Contrastive Pose-EMG Pre-training (CPEP):

   • Function: Freezes the pose encoder \(\mathcal{E}_p\) while training the EMG encoder together with a projection head \(h\).
   • Mechanism: EMG embeddings are computed as \(u_i = h(\mathcal{E}_x(x_i))_{[\text{CLS}]}\) and pose embeddings as \(v_i = (\mathcal{E}_p(p_i))_{[\text{CLS}]}\). After \(\ell_2\) normalization, symmetric InfoNCE is applied: \(\mathcal{L}_{\text{CPEP}} = \frac{1}{2N}\sum_{i} [-\log\frac{\exp(s_{ii})}{\sum_j\exp(s_{ij})} - \log\frac{\exp(s_{ii})}{\sum_j\exp(s_{ji})}]\), where \(s_{ij} = \tilde{u}_i^\top\tilde{v}_j / \tau\).
   • Design Motivation: Freezing the pose encoder is critical—jointly updating both encoders degrades pose representation quality and leads to training divergence, as confirmed by ablation experiments.

3. Zero-Shot Classification Protocol:

   • Function: \(k\)-nearest neighbor voting in the embedding space.
   • Mechanism: Pose embeddings are pre-computed offline; for each EMG query, the top-\(k\) (\(k=10\)) nearest pose embeddings are retrieved and a majority vote determines the predicted label: \(\hat{y}_j = \text{mode}\{y(p) | p \in \mathcal{R}_j\}\).
   • Design Motivation: Zero-shot performance validates that the EMG representations have internalized the structural information of hand poses.

Loss & Training

Three-stage training pipeline: EMG/Pose-MAE pre-training for 100 epochs each → CPEP contrastive pre-training for 100 epochs (batch size 256, learnable temperature \(\tau\) initialized at 0.02) → linear probing. Training is conducted on 4× V100 GPUs and takes approximately 4.5 hours per model.

Key Experimental Results

Main Results (Gesture Classification Accuracy)

Method	LP In-Dist.	LP Unseen	ZS In-Dist.	ZS Unseen
emg2pose (baseline)	0.647	0.312	-	-
EMG-MAE	~0.55	~0.30	-	-
PoseT (supervised)	~0.60	~0.35	-	-
CPEP	0.782	0.536	0.757	0.481
Pose-MAE (upper bound)	~0.85	~0.65	-	-

Ablation Study

Configuration	LP In-Dist.	ZS In-Dist.	LP Unseen	ZS Unseen
EMG encoder Frozen	0.372	0.344	0.326	0.298
EMG encoder RandInit	0.748	0.701	0.479	0.454
AvgPool	0.761	0.711	0.518	0.454
CPEP (full)	0.782	0.757	0.536	0.481

Key Findings

• MAE pre-training initialization is critical: random initialization leads to slower convergence and lower accuracy, and jointly training both encoders fails to converge.
• [CLS] token outperforms AvgPool, indicating that global context is more effective for gesture recognition.
• Freezing the EMG encoder yields drastically worse performance (0.372 vs. 0.782), confirming that fine-tuning the EMG encoder is necessary.
• Longer EMG patches degrade performance, underscoring the need for fine-grained temporal modeling.

Highlights & Insights

• First zero-shot gesture recognition framework for EMG: The zero-shot results surpass the baseline's linear probing performance (0.481 vs. 0.312 on unseen gestures), demonstrating that contrastive pre-training yields representations with genuine generalization capability.
• The strong-modality-anchors-weak-modality paradigm is transferable to analogous settings such as IMU–video alignment and EEG–behavior alignment.

Limitations & Future Work

• Validation is limited to a single dataset (emg2pose); generalizability to other EMG acquisition devices and protocols remains untested.
• The gesture vocabulary is small (4+4 classes), whereas practical applications require recognition of dozens to hundreds of gestures.
• No comparison is made against advanced contrastive learning methods such as SigLIP or CLAP.
• As a workshop paper, the experimental scale is limited and statistical significance is not reported.
• Online adaptation and few-shot fine-tuning scenarios are not explored.
• Robustness to inter-subject EMG signal variability is insufficiently analyzed.

Related Work & Insights

• vs. emg2pose: Supervised pose regression offers limited generalization; CPEP's contrastive alignment produces structured embeddings that support zero-shot retrieval.
• vs. CLIP: CPEP adopts the cross-modal contrastive learning paradigm but introduces key adaptations—pre-training encoders to reduce data requirements and freezing the strong modality encoder.
• vs. NeuroPose/Vemg2pose: These baselines also employ Transformer architectures but are trained with supervised regression objectives, yielding embeddings of insufficient quality for retrieval-based classification.

Rating

• Novelty: ⭐⭐⭐⭐ First application of CLIP-style contrastive pre-training to EMG–pose alignment for zero-shot gesture recognition.
• Experimental Thoroughness: ⭐⭐⭐ Workshop paper; single dataset; limited gesture vocabulary.
• Writing Quality: ⭐⭐⭐⭐ Problem formulation is clear and method description is concise.
• Value: ⭐⭐⭐⭐ Opens a new direction for zero-shot EMG-based gesture recognition.

Related Papers

• [AAAI 2026] New Synthetic Goldmine: Hand Joint Angle-Driven EMG Data Generation Framework for Micro-Gesture Recognition
• [ICCV 2025] OpenAnimals: Revisiting Person Re-Identification for Animals Towards Better Generalization
• [NeurIPS 2025] PandaPose: 3D Human Pose Lifting from a Single Image via Propagating 2D Pose Prior to 3D Anchor Space
• [NeurIPS 2025] RAPTR: Radar-Based 3D Pose Estimation Using Transformer
• [NeurIPS 2025] Cycle-Sync: Robust Global Camera Pose Estimation through Enhanced Cycle-Consistent Synchronization