npu-detection

What problem does it solve?

Deploying YOLO-based object and defect detection models to NPU hardware often leads to significant accuracy loss from improper quantization and requires manual tuning of inference parameters to meet latency and throughput requirements for edge AI camera systems.

Core Features & Use Cases

• End-to-End NPU Deployment Pipeline: Covers full workflow from ONNX model validation and input dimension checks to quantization, accuracy verification, and post-processing tuning for RKNN, TFLite, and TensorRT runtimes.
• Built-In Accuracy Guardrails: Enforces a maximum 2% mAP drop after quantization, with fallback options including mixed-precision quantization and Quantization-Aware Training if accuracy thresholds are not met.
• Use Case: Edge AI development teams can use this skill to deploy defect detection models for industrial or security cameras to NPU-powered devices, ensuring real-time inference performance with validated accuracy and configurable NMS parameters for edge cases like small or overlapping objects.

Quick Start

Use the npu-detection skill to deploy your trained YOLO object detection model to the target NPU hardware, run quantization with the provided calibration dataset, and verify it meets the required mAP and latency thresholds.