SqueakPose Studio: An end-to-end platform for pose estimation and real-time edge-AI deployment

Area of Science:

• Behavioral neuroscience
• Computational biology
• Machine learning

Background:

• Accurate pose estimation is crucial for quantitative behavior analysis.
• Current deep learning tools often require complex, offline workflows and specialized hardware.
• Real-time deployment of pose estimation models is challenging due to software fragmentation and hardware limitations.

Purpose of the Study:

• To present an integrated software-hardware ecosystem for pose estimation.
• To enable seamless transition from dataset creation to real-time deployment on edge devices.
• To overcome limitations of existing offline-centric pose estimation tools.

Main Methods:

• Developed SqueakPose Studio for dataset creation, labeling, training, and offline inference.
• Utilized modern object-detection architectures for efficient end-to-end training and inference.
• Integrated SqueakView and MouseHouse for real-time deployment on embedded edge-computing hardware.

Main Results:

• The system supports efficient pose estimation across CPUs, GPUs, and Apple Silicon.
• Enabled real-time model deployment, video capture, and sensor logging on edge devices.
• Ensured consistency between offline analysis and real-time deployment through a shared data format and deterministic timing.

Conclusions:

• The SqueakPose ecosystem provides a unified platform for both offline and real-time pose estimation.
• Facilitates behavior analysis on embedded edge-computing devices without reliance on workstation-grade hardware.
• Advances the accessibility and application of deep learning for quantitative behavior analysis in experimental settings.