Skip to content

README.md

Latest commit

 

History

History
99 lines (73 loc) · 4.13 KB

README.md

File metadata and controls

99 lines (73 loc) · 4.13 KB
Raw

Face Detection + Face Recognition Pipeline

This example demonstrates Face Detection / Recognition utilizing off-the-shelf YoloV8n-Face + FaceNet models co-mapped and running on MemryX accelerator. It implements a face recognition pipeline where:

  1. Face boxes and keypoints are detected an image.
  2. Faces are extracted from the image using bounding boxes and keypoints.
  3. Extracted faces are embedded using the face recognition model.

The implementation includes MXFace, which emulates a Queue structure, making it easily integrable into a realtime application. The following demo code utilizes MXFace to distinguish the identity between two images.

Same Image Different Image

Overview

Property Details
Model YoloV8n-Face, FaceNet
Model Type Face Detection + Recognition
Framework Onnx + Keras
Model Source YoloV8n-Face, FaceNet
Pre-compiled DFP Download here
Output Face bounding box + keypoints + embedding
OS Linux
License GPL

Requirements

Before running the application, ensure that OpenCV and Kagglehub re installed. You can install them using the following commands:

pip install -r requirements.txt

Run the Application

Step 1: Download Pre-compiled DFP

To download and unzip the precompiled DFPs, use the following commands:

wget https://developer.memryx.com/example_files/2p2/face_recognition.zip
mkdir -p models
unzip face_recognition.zip -d models
(Optional) Download and compile the model yourself If you prefer, you can download and compile the models rather than using the precompiled model. Download the pre-trained yolov8n and FaceNet models: 
wget https://developer.memryx.com/example_files/face_recognition_original.zip
unzip face_recognition_original.zip -d models

You can now use the MemryX Neural Compiler to compile the models and generate the DFP file required by the accelerator:

cd models/ 
mx_nc -v -m Facenet.h5 yolov8n-face_crop.onnx --dfp_fname yolov8n_facenet.dfp

Step 2: Run the Script

With the compiled model, you can now use the MXA to perform face recognition. Run the following script to see face recognition in action:

cd src/
python3 run.py

By default, the script will download a sample dataset of images from hit comedy T.V. series Friends and compare two sample images. Additionally, you can specify two images to distinguish:

python3 run.py --image_path path1 path2

Third-Party Licenses

This project uses third-party software, models, and libraries. Below are the details of the licenses for these dependencies:

Summary

This example impelments a Face Detection + Recognition pipeline, utilizing off-the shelf yolov8 + facenet models.