face_amnesia

Detect and retrieve the face of a specific person in image and video files using pre-trained models.

Requirements

Language

• Python (>= 3.5)

Dependencies

• NumPy (http://www.numpy.org/)
  • basic numerical computations
• scikit-learn (https://scikit-learn.org/stable/)
  • PCA
• OpenCV (https://opencv.org/)
  • basic image and video processing
• dlib (http://dlib.net/)
  • face detection and recognition models

Install

Use pre-compiled libraries

Run provided install script to use pre-compiled versions of OpenCV and dlib (without GPU support) available via pip. Note that corresponding pre-built OpenCV package is unofficial but works fine. See comments in install script for further details.

Use Dockerfile

Run provided Dockerfile to compile OpenCV and dlib by hand using the official Ubuntu 20.04 image.

cd /path/to/face_amnesia/
docker build -t face_amnesia .
docker run -it face_amnesia

Note that it is possible to enable hardware-dependent optimization such as AVX instructions or CUDA support. See comments in Dockerfile for further details.

Testing

Run provided test script to execute unit and integration tests.

Usage

Command-line interfaces for data point creation and retrieval available in project's root directory. Data points and corresponding source files are stored in and served from a separate face amnesia folder in current user's home directory which may be identical to project's root directory.

Data point creation

Create face description vectors from media file (image, video) or whole directory.

python3 create.py /path/to/media/file/or/directory/

Available options:

python3 create.py --help

Data point retrieval

Retrieve previously stored data points corresponding to faces similar to given one.

python3 retrieve.py /path/to/query/media/file/

Available options:

python3 retrieve.py --help

Further reading

• Mayur Datar, Nicole Immorlica, et. al.: "Locality-Sensitive Hashing Scheme Based on p-Stable Distributions" (2004)
• Chris Biemann: "Chinese Whispers - an Efficient Graph Clustering Algorithm and its Application to Natural Language Processing Problems" (2006)

Further tools

• Adam Geitgey: face_recognition (https://github.com/ageitgey/face_recognition)

Troubleshooting

Compiling dlib by hand is usually straightforward. Building OpenCV from source, however, may need some adjustments. The official guides on this topic are useful and can be found on the web:

• http://dlib.net/compile.html
• https://docs.opencv.org/4.12.0/d7/d9f/tutorial_linux_install.html

The following paragraphs will provide additional information on compilation problems which may occur on some architectures.

Compilation: CBLAS / LAPACK headers could not be found

Possible workaround:

sudo cp /usr/include/lapacke*.h /usr/include/openblas/

Compilation: GCC version 8 is not supported

Possible workaround:

export CC=/path/to/gcc-7/compiler

Building dlib's Python bindings with CUDA support needs an additional argument to setup.py call:

--set CUDA_HOST_COMPILER=/path/to/gcc-7/compiler