Facial Landmark Detection Quantized

Introduction

The Facial-Landmark-Detection-Quantized model by Qualcomm is designed for real-time 3D facial landmark detection. It is optimized for mobile and edge deployment, supporting multiple formats including TFLite, QNN, and ONNX. This model is specifically tailored for use on Qualcomm devices and targets efficient mobile deployment.

Architecture

The model is based on pose estimation techniques. It accepts an input resolution of 128x128 and contains 5.424 million parameters, resulting in a model size of 5.314 MB. The architecture is optimized to run efficiently on various Qualcomm chipsets, utilizing INT8 precision for computations primarily on NPUs.

Training

The model is trained using methodologies suitable for quantization, allowing it to perform efficiently on devices with limited computational resources. The training process focuses on optimizing the model for inference speed and memory usage on Snapdragon processors.

Guide: Running Locally

1. Installation: Install the model package using pip:

   pip install "qai-hub-models[facemap_3dmm_quantized]"
   

2. Configuration: Configure your environment with your Qualcomm AI Hub API token:

   qai-hub configure --api_token YOUR_API_TOKEN
   

3. Demo Execution: Run a simple demo to test the model:

   python -m qai_hub_models.models.facemap_3dmm_quantized.demo
   

   For Jupyter or Colab, use:

   %run -m qai_hub_models.models.facemap_3dmm_quantized.demo
   

4. Cloud Execution: Execute the model on a cloud-hosted Qualcomm device:

   python -m qai_hub_models.models.facemap_3dmm_quantized.export
   

5. Deploying to Android: Follow TensorFlow Lite quickstart guide for .tflite deployment. For QNN, refer to Qualcomm's sample app documentation.

Suggested Cloud GPUs: Consider using instances with Qualcomm Snapdragon support to maximize performance.

License

The original model implementation is licensed under BSD-3-Clause, available here. The compiled assets for on-device deployment are covered under Qualcomm AI Hub's proprietary license, accessible here.