Creates 256 cropped, high quality facesets from a batch of mp4 videos end to end. Modified code from https://github.com/Tyrannosaurus1234/GetFaces by Tyrannosaurus1234 adjusted for batch converting and extra quality scanning of the final faceset by removing photos below a minimum size and double checking there is only one face in the photo
$python3 demo.py -i '/PATH/TO/PASSPORT/STYLE/PHOTO/OF/TARGET/FACE.jpg' -v '/PATH/TO/DIR/OF/MP4s'
-t = Tolerance of face detection, lower is stricter (0.1-1.0) Default = 0.6
-f = Amount of frames per second to scan. Default = 25
-n = Number photos of the face from each video. Default = 1000 (Set higher if you have less vids)
-s = Minimum KB size of images to keep in the faceset. Default = 32
Running the script creates three files in your directory. One for the faceset. One for the scanned mp4s. One for the small faces.
The batch processing of the faceset can be stopped or started anytime without a problem, aside from maybe it creating duplicates if it stops in the middle of a video. Just go in your faceset folder and remove photos from the half processed video when you start it again.
The script can read videos other than mp4s. Just do a find and replace in the script to your extension, like mp4 to mkv.
Best to use mp4s ripped from YouTube. Use YouTube video downloader sites or apps to grab dozen(s) of 1080p videos starring
the target face.
numpy
opencv-python >= 3.31
face_recognition
face_recognition requires a package called dlib. If you have a GPU you must install dlib in a specific way in order to get it to use your GPU ...Which doesn't seem to be listed in the documentation but doing so is as follows,
git clone https://github.com/davisking/dlib.git
cd dlib
mkdir build; cd build; cmake .. -DDLIB_USE_CUDA=0 -DUSE_AVX_INSTRUCTIONS=1; cmake --build .
cd ..
python3 setup.py install --yes USE_AVX_INSTRUCTIONS --no DLIB_USE_CUDA
Also included is a dulpicate image finder you might find usefull on batches of Google image downloads or videos with overlapping footage of your desired face. Maybe useful for ensuring you made a clean dataset. Be mindful it might think high frame rate ripped stills are duplicates when you want keep those subtly different frames, so it needs your supervision.
This doesn't make the alignments file. This is for building and adding on to your faceset. The program that makes the alignment file, rotates, and crops the images will find faces but this will find YOUR desired face and this will speed up the process.
5000 is a good number if you have various facial expressions in various lighting.
You can do 50,000 or 100,000 and beyond.