Bob interface for mtcnn face and landmark detection

This package binds the face and landmark detection from the paper:

@ARTICLE{7553523,

author={K. Zhang and Z. Zhang and Z. Li and Y. Qiao}, journal={IEEE Signal Processing Letters}, title={Joint Face Detection and Alignment Using Multitask Cascaded Convolutional Networks}, year={2016}, volume={23}, number={10}, pages={1499-1503}, keywords={Benchmark testing;Computer architecture;Convolution;Detectors;Face;Face detection;Training;Cascaded convolutional neural network (CNN);face alignment;face detection}, doi={10.1109/LSP.2016.2603342}, ISSN={1070-9908}, month={Oct},}

User guide

Face Detection and Landmark detection

The most simple face detection task is to detect a single face in an image. This task can be achieved using a single command:

>>> import bob.ip.mtcnn
>>> import bob.io.base
>>> import bob.io.base.test_utils
>>> color_image = bob.io.base.load(bob.io.base.test_utils.datafile('testimage.jpg', 'bob.ip.facedetect'))
>>> bounding_box, landmarks = bob.ip.mtcnn.FaceDetector().detect_single_face(color_image)
>>> print(bounding_box.topleft)
(64, 77)

(Source code, png, hires.png, pdf)

_images/plot_single_faces.png

Multiple Face Detection

The detection of multiple faces can be achieved with a single command:

>>> import bob.ip.mtcnn
>>> import bob.io.base
>>> import bob.io.base.test_utils
>>> color_image = bob.io.base.load(bob.io.base.test_utils.datafile('multiple-faces.jpg', 'bob.ip.mtcnn'))
>>> bounding_box, landmarks = bob.ip.mtcnn.FaceDetector().detect_all_faces(color_image)
>>> print ((bounding_box[0].topleft, bounding_box[0].bottomright))
((28, 189), (102, 244))

(Source code, png, hires.png, pdf)

_images/plot_multiple_faces.png

Landmark detection

The detection of landmarks can be done as the following:

>>> import bob.ip.mtcnn
>>> import bob.io.base
>>> import bob.io.base.test_utils
>>> color_image = bob.io.base.load(bob.io.base.test_utils.datafile('testimage.jpg', 'bob.ip.facedetect'))
>>> bounding_box, landmarks = bob.ip.mtcnn.FaceDetector().detect_single_face(color_image)
>>> print (landmarks['leye'])
(174, 219)

Face genometric normalization

The detection of landmarks can be done as the following:

>>> import bob.ip.mtcnn
>>> import bob.io.base
>>> import bob.io.base.test_utils
>>> color_image = bob.io.base.load(bob.io.base.test_utils.datafile('testimage.jpg', 'bob.ip.facedetect'))
>>> color_image_norm = bob.ip.mtcnn.FaceDetector().detect_crop(color_image, final_image_size=(224, 224))
>>> color_image_norm_align = bob.ip.mtcnn.FaceDetector().detect_crop_align(color_image, final_image_size=(224, 224))

(Source code, png, hires.png, pdf)

_images/plot_align_faces.png