Advantech-EIOT/qualcomm-lightweight_face_detection

qualcomm-lightweight-face-detection

• Original model: Lightweight-Face-Detection

This repository contains the Lightweight-Face-Detection model optimized for Qualcomm hardware using Qualcomm® AI Engine Direct (QNN).

It is designed for high-performance, real-time face detection inference on edge devices powered by Qualcomm Snapdragon platforms, enabling efficient on-device AI capabilities with low latency and reduced power consumption.

Model Details

• Developed by: Advantech-EIOT / Qualcomm
• Architecture: Lightweight Face Detection
• Task: Face Detection
• Precision: Quantized (w8a8) for NPU optimization
• Input Resolution: 480 x 640
• Optimization: Qualcomm® AI Stack / QNN SDK

Hardware Compatibility

This model is highly optimized for Advantech Edge AI platforms powered by Qualcomm processors:

• Linux: Dragonwing® Platforms (e.g. Dragonwing® IQ-9075)

Limitations and Disclaimer

Lightweight-Face-Detection is a powerful real-time face detection model, but it may exhibit limitations depending on the environment and deployment context.

• Accuracy: The model's accuracy, especially for small faces, extreme head poses, or in low-light/backlit conditions, may be impacted by the quantization (w8a8). Users should validate outputs for critical applications or safety-critical systems.
• Usage: Please refer to the BSD-3-CLAUSE for usage restrictions and acceptable use policies.
• Edge Optimization: Inference performance (FPS) and bounding box precision may vary depending on the specific hardware configuration, camera ISP pipelines, and thermal constraints of the edge device.