Pose2Muscle: Structured Spatio-Temporal Decoding for Discrete Muscle Activity Estimation from Human Pose
Yuepeng Chen, Jiehong Shi, Kaili Zheng, Boyi Zhang, Chenyi Guo, Ji Wu, Xiangling Fu
cs.CV
Quick verdict
Pose-to-discrete-muscle-state estimation with a DAG decoder and a new pose-sEMG dataset; sound, but cross-subject gains are small and lack significance tests.
Summary
The paper proposes Pose2Muscle, a framework that predicts discrete muscle activity states directly from 3D pose sequences, removing the need for sEMG at inference. Continuous sEMG (16 channels over 8 bilateral muscle groups, Delsys Trigno, 1259 Hz) is tokenized into 13 ordinal states per muscle via handcrafted features and K-means, following the authors' prior work. The model combines a multi-scale spatio-temporal attention encoder (intra-/inter-window temporal and intra-/inter-limb spatial attention, adapted from SkateFormer) with a DAG-based decoder (adapted from DA-Transformer) that keeps lambda candidate state hypotheses per time step, trains via marginalized path likelihood, and decodes the most probable trajectory. The authors contribute PoseEMG-43 (2,992 instances, 43 actions, 14 participants; OptiTrack pose at 120 Hz converted to 25-joint NTU skeletons) plus a small MEQAD set of correctly/incorrectly performed exercises. Against retrieval, ST-GCN, PoseFormer, MotionBERT, SkateFormer, and MIA-Model baselines, Pose2Muscle reports Adj-ACC 86.36%/PCC 0.8821 (random split) and 63.97%/0.6795 (5-fold subject-level CV); ablations show both modules help. Main concerns: the random split does not control subject/action overlap (retrieval alone reaches PCC 0.7757), cross-subject gains are small and untested for significance, the dataset is unreleased, and inputs are laboratory mocap rather than video.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
4 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Discretizing sEMG into 13 ordinal muscle-state tokens provides more stable and interpretable supervision than direct continuous sEMG regression.
- Combining multi-scale spatio-temporal attention with a DAG-based multi-hypothesis decoder improves pose-to-muscle-state prediction over retrieval and strong motion-representation baselines on PoseEMG-43.
- Pose2Muscle generalizes to unseen subjects, achieving Adj-ACC 63.97% and PCC 0.6795 under five-fold subject-level cross-validation, with the DAG decoder contributing most in this setting.
- Models trained only on PoseEMG-43 transfer without fine-tuning to the external MEQAD set, distinguishing correct from faulty exercise executions such as deltoid compensation in biceps curls.
Strengths
Well-motivated reformulation of the target space
Regressing non-stationary raw sEMG is replaced by predicting 13 ordinal muscle states tokenized via K-means on training-split features only (Sec 4.1.3 avoids label leakage), and adjacent-level accuracy respects the ordinal structure. This consistently yields the best results against all six baselines in both splits (Table 2: Adj-ACC 86.36% random, 63.97% subject-level).
New synchronized pose-sEMG resource
PoseEMG-43 provides 2,992 instances across 43 daily/exercise actions from 14 participants, pairing OptiTrack mocap (120 Hz, Mosh++-processed 25-joint NTU-compatible skeletons) with 16-channel Delsys Trigno sEMG over 8 bilateral muscle groups. NTU formatting aids reuse, and the expert-designed MEQAD set adds an execution-quality external test.
Novel decoder adaptation with supporting ablations
Multi-protocol, multi-angle evaluation
Weaknesses
Random split may leak subject/action context
Small cross-subject gains without statistical testing
Self-referential target space and metric caveats
Reproducibility and reporting-rigor gaps
Mocap-only input limits the practical motivation
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Can muscle activity be estimated from human pose without EMG sensors?
This paper argues yes, approximately. It converts surface EMG into 13 discrete activity states per muscle and trains a model to predict those states from 3D pose sequences alone. On its new PoseEMG-43 dataset it reports 86.4% adjacent-level accuracy (predictions within one state level) in same-distribution testing, dropping to 64.0% for unseen people. This supports coarse muscle-aware feedback for fitness and rehabilitation, but it is not a clinical sEMG replacement, and the study used laboratory motion capture rather than ordinary video.
What is the PoseEMG-43 dataset?
A synchronized pose-EMG dataset with 2,992 movement instances covering 43 daily-life and exercise actions (walking, jogging, squats, jumps, arm exercises, and more) performed by 14 adults aged 21-36. Pose was captured with an OptiTrack system at 120 Hz and converted to standard 25-joint NTU-RGB+D-style skeletons; muscle activity was recorded with 16 Delsys Trigno wireless sEMG channels covering eight bilateral muscle groups (deltoid, biceps brachii, rectus abdominis, rectus femoris, tibialis anterior, middle trapezius, biceps femoris, lateral gastrocnemius). The authors state it will be released publicly after paper acceptance.
How accurate is Pose2Muscle?
Under a random train/test split it reaches 86.36% adjacent-level accuracy and 43.54% exact top-1 accuracy across 13 states, with a Pearson correlation of 0.8821 between predicted and true state sequences. Under five-fold cross-validation on unseen subjects, performance drops to 63.97% adjacent accuracy and 0.6795 PCC, showing substantial inter-subject variability in the pose-to-muscle relationship. Most remaining errors are within one or two state levels.
Can this method detect incorrect exercise form?
Preliminarily, yes. On a small external dataset (MEQAD) of correctly and incorrectly performed high knees, squats, and biceps curls, demonstrated by a certified strength coach, the model transferred without any fine-tuning and clearly beat a pose-similarity retrieval baseline (59.7% vs 43.3% adjacent accuracy). It distinguished compensations such as the anterior deltoid taking over in faulty biceps curls. The evaluation set is small, so this capability remains an early demonstration rather than a validated rehabilitation tool.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures (confusion matrices, error distributions, lambda-sensitivity curves, qualitative muscle-state heatmaps) could not be visually inspected; assessments of them rely solely on the authors' textual descriptions.
- MEQAD is described qualitatively; per-action sample counts, annotation procedure, and synchronization accuracy between sEMG (1259 Hz) and pose (120 Hz) streams are not quantified in the extracted text.
- The extraction is a preprint formatted for Elsevier and appears complete (23 pages, appendix, references), but appendix A.1 is brief, so data-collection details such as sensor placement standardization across participants could not be verified.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.