giacomopiccinini/facedetection

Face detection using dlib, for images, videos and stream. Blur faces and save results!

Face Detection Using dlib

This repository contains the code implementing a convenient high-level API for face detection using dlib. Admissible files are: images, videos and stream from computer's webcam. In addition to simply detecting the face, it is possible to:

• Blur the detected faces;
• Blur the background;
• Recognise people whose faces are detected (if examples are provided, of course);
• Show detection results as the script runs;
• Save detections to disk.

Installing dlib

dlib is not particulatly friendly to install, in particular if one wants to enable GPU acceleration. Since dlib uses both CUDA and cuDNN, you need to have both installed. After that, create a venv and then proceed as follows:

$ git clone https://github.com/davisking/dlib.git
$ cd dlib
$ mkdir build
$ cd build
$ cmake .. -DDLIB_USE_CUDA=1 -DUSE_AVX_INSTRUCTIONS=1 
$ cmake --build .
$ cd ..
$ python setup.py install --set DLIB_USE_CUDA=1

After that you should have installed dlib with CUDA. To check that this is the case, you could run

import dlib
print(dlib.cuda.get_num_devices())
print(dlib.DLIB_USE_CUDA)

The first print statement indicates if CUDA-capable devices are actually recognised. However, even if recognised, a CUDA device might not be activate for use with dlib. The second print statement precisely checks that this is the case. Obviously, it is not possible to forcefully set dlib.DLIB_USE_CUDA=True.

Launching the script

To simply detect faces in a particular file, run

python main.py --source <file>

If, instead of a file, you pass a directory, e.g. python main.py --source <directory> then all files in that directory (and subdirectories, too) are recursively processed. Conversely, if you would like to use your webcam, then

python main.py --source webcam

In no source flag is passed, the script will automatically read from the Input directory I have provided.

Options

• --model <path_to_model>: the model to be loaded for inference. By default it looks in the Model directory, but you can change that at your will.
• --upsample <integer>: if the face is too small, it might not be detected at a first try. To remedy that, you could pass an integer greater than 1: this will upscale the image making it easier to detect small faces.
• --show: when this flag is passed, results of detection will be showed at inference time. Notice: this is the default when webcam is active!
• --save: when this flag is passed, results of the dection will be store in the Output directory. The name is the same as that of the original file.
• --blur <integer>: if <integer> is greater than zero, the face is blurred. The higher the integer, the more violenet the blur is. If <integer> is smaller than zero, the background is blurred.
• --recognize <path_to_examples_directory>: it is possible to recognize faces in the picture, provided few pictures of the relevant people are passed. The example directoryis to be organised as follows:

Examples_Directory
│
└── <name_1>
│      │  
│      │ <picture_1>
│      │ <picture_2>
│      │ ...
└── <name_2>
│      │  
│      │ <picture_3>
│      │ <picture_4>
│      │ ...
...
                  

Issues

Due to dlib implementation, you will likely incur in malloc issues. For some reason, the GPU memory will fill up extremely quickly and you will run out of memory. A simple solution could be to resize images. For instance, you could use Image Magick and run

mogrify -resize 256x256 ./*.png

to resize to 256x256 all .png images in subdirectories of the directory you are launching the script from.v