Reproduce “Re-evaluating Position and Velocity Decoding for Hand Pose Estimation with Surface Electromyography” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Nima Hadidi, Johannes Lee, Ebrahim Feghhi, Michael Yuan, Jonathan C. Kao

Researchers re-evaluated the emg2pose benchmark's conclusion regarding sEMG hand pose estimation, finding that position decoding, when stably optimized, achieves superior accuracy for the Tracking task compared to velocity decoding. A lightweight causal filter further enables position decoding to match or exceed velocity decoding's smoothness, offering a revised approach for real-time applications.

Key insights

  • Stable training for LSTM-based position decoding models critically depends on proper tuning of the decoder output scalar; an inappropriate fixed value can lead to optimization collapse.
  • Position decoding is inherently more robust to long-term error accumulation (better low-frequency accuracy) for continuous pose prediction, while its higher-frequency jitter can be effectively mitigated with post-processing.
  • Multi-task training, specifically combining Tracking and Regression objectives, significantly enhances the temporal coherence and overall performance of models on the more challenging Regression task.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.