This paper is accepted in Basic models of brain and body assembly in neurips 2025.
Hand segmentation using high-quality structured data such as videos, images, and hand bones is a well-studied problem in computer vision. Leveraging Energy Potential Smaller, more expensive Biosignigners, e.g. In this paper, we show that learning representations from weak data such as data aligned with those from structured, high-quality data can improve the quality of the representation and enable zero-tusher classification. Specifically, we propose a pose-EMG pre-training framework for synchronizing EMG representations with pose, where we learn an EMG encoder that produces high-quality and informative representations. We test the performance of our model's touch phases using linear reshop and zero-shot setups. Our models derive from the EMG2Forum models up to 21% in the classification of the moving action and 72% in the invisible (output-diffusion) classification of the body.
- † University of California
- ** Work Done while at Apple
