npu-pose

What problem does it solve?

Deploying human pose and gesture estimation models to NPU hardware often leads to reduced keypoint accuracy due to quantization, making it difficult to meet performance requirements for real-time edge applications like AI camera body tracking.

Core Features & Use Cases

• Model Selection Guidance: Choose between top-down and bottom-up pose estimation architectures, with recommendations for MoveNet, HRNet, and MediaPipe Pose models.
• NPU Quantization Tuning: Optimize model performance with INT8 quantization, with automatic fallback to FP16 if Object Keypoint Similarity (OKS) scores drop below acceptable thresholds.
• Accuracy Validation: Verify deployment performance using COCO keypoint validation datasets, with per-joint accuracy analysis and temporal jitter checks for video sequences.
• Use Case: Ideal for embedded AI camera systems that require low-latency, accurate body tracking for applications like retail analytics, security monitoring, and fitness tracking.

Quick Start

Use the npu-pose skill to deploy your selected pose estimation model to the target NPU and confirm it meets the required [email protected] accuracy and latency thresholds.