This package implements parts of Google®'s MediaPipe models in pure Python (with a little help from Numpy and PIL) without Protobuf
graphs and with minimal dependencies (just TF Lite and Pillow).
The package provides the following models:
- Face Detection
- Face Landmark Detection
- Iris Landmark Detection
- Iris recoloring example
The package doesn't use the graph approach implemented by MediaPipe and is therefore not as flexible. It is, however, somewhat easier to use and understand and more accessible to recreational programming and experimenting with the pretrained ML models than the rather complex MediaPipe framework.
Here's how face detection works and an image like shown above can be produced:
from fdlite import FaceDetection, FaceDetectionModel
from fdlite.render import Colors, detections_to_render_data, render_to_image
from PIL import Image
image = Image.open('group.jpg')
detect_faces = FaceDetection(model_type=FaceDetectionModel.BACK_CAMERA)
faces = detect_faces(image)
if not len(faces):
print('no faces detected :(')
else:
render_data = detections_to_render_data(faces, bounds_color=Colors.GREEN)
render_to_image(render_data, image).show()
While this example isn't that much simpler than the MediaPipe equivalent, some models (e.g. iris detection) aren't available in the Python API.
Note that the package ships with five models:
FaceDetectionModel.FRONT_CAMERA
- a smaller model optimised for selfies and close-up portraits; this is the default model usedFaceDetectionModel.BACK_CAMERA
- a larger model suitable for group images and wider shots with smaller facesFaceDetectionModel.SHORT
- a model best suited for short range images, i.e. faces are within 2 metres from the cameraFaceDetectionModel.FULL
- a model best suited for mid range images, i.e. faces are within 5 metres from the cameraFaceDetectionModel.FULL_SPARSE
- a model best suited for mid range images, i.e. faces are within 5 metres from the camera
The FaceDetectionModel.FULL
and FaceDetectionModel.FULL_SPARSE
models are
equivalent in terms of detection quality. They differ in that the full model
is a dense model whereas the sparse model runs up to 30% faster on CPUs. On a
GPU, both models exhibit similar runtime performance. In addition, the dense
full model has slightly better Recall,
whereas the sparse model features a higher Precision.
If you don't know whether the image is a close-up portrait or you get no
detections with the default model, try using the BACK_CAMERA
-model instead.
The latest release version is available in PyPI and can be installed via:
pip install -U face-detection-tflite
The package can be also installed from source by navigating to the folder
containing setup.py
and running
pip install .
from a shell or command prompt.