face-comparison

AI Face comparison using FaceNet, compare two photos and see if they are the same person.

Installation

pip install face-compare

Usage

Use compare_faces.py to compare two images of people to see if they are the same person.

compare_faces.py --image-one /path/to/image_one.png --image-two /path/to/image_two.png

Optionally output the cropped image output to a directory (useful for inspecting input to AI model)

compare_faces.py --image-one /path/to/image_one.png --image-two /path/to/image_two.png -s /path/to/outputs/

Steps Involved

1. A cascade classifier is used to detect the face within the input images.
2. The bounding box of this segmentation is then used to crop the images, and fed into the AI model.
3. The FaceNet model then calculates the image embeddings for the two cropped images.
4. Finally the second embedding is subtracted from the first, and the Euclidean norm of that vector is calculated.
5. A threshold of 0.7 is used to determine whether they are the same person or not.

Known Issues

CPU Only runtime issue

If you are trying to run the module without a suitable GPU, you may run into the following error message:

tensorflow.python.framework.errors_impl.InvalidArgumentError:  Default MaxPoolingOp only supports NHWC on device type CPU

To fix this issue with Intel CPU architecture, you can install the TensorFlow Intel Optimization package via

pip install intel-tensorflow

References

This module uses the AI model FaceNet, which can be found here, and the journal article here.