Implementation of the MTCNN face detector for TensorFlow in Python3.4+. It is written from scratch, using as a reference the implementation of MTCNN from David Sandberg (FaceNet's MTCNN) in Facenet. It is based on the paper Zhang, K et al. (2016) [ZHANG2016].
Currently it is only supported Python3.4 onwards. It can be installed through pip:
$ pip3 install mtcnnThis implementation requires OpenCV>=4.1 and Tensorflow>=1.12.1 installed in the system, with bindings for Python3.
They can be installed through pip (if pip version >= 9.0.1):
$ pip3 install tensorflow==1.12.1 opencv-contrib-python==4.1.0or compiled directly from sources (OpenCV4, Tensorflow).
Note that a tensorflow-gpu version can be used instead if a GPU device is available on the system, which will speedup the results. It can be installed with pip:
$ pip3 install tensorflow-gpu\>=1.12.0The following example illustrates the ease of use of this package:
>>> from mtcnn.mtcnn import MTCNN
>>> import cv2
>>>
>>> img = cv2.imread("ivan.jpg")
>>> detector = MTCNN()
>>> print(detector.detect_faces(img))
[{'box': [277, 90, 48, 63], 'keypoints': {'nose': (303, 131), 'mouth_right': (313, 141), 'right_eye': (314, 114), 'left_eye': (291, 117), 'mouth_left': (296, 143)}, 'confidence': 0.99851983785629272}]The detector returns a list of JSON objects. Each JSON object contains three main keys: 'box', 'confidence' and 'keypoints':
- The bounding box is formatted as [x, y, width, height] under the key 'box'.
- The confidence is the probability for a bounding box to be matching a face.
- The keypoints are formatted into a JSON object with the keys 'left_eye', 'right_eye', 'nose', 'mouth_left', 'mouth_right'. Each keypoint is identified by a pixel position (x, y).
Another good example of usage can be found in the file "example.py." located in the root of this repository.
The following tables shows the benchmark of this mtcnn implementation running on an Intel i7-3612QM CPU @ 2.10GHz, with a CPU-based Tensorflow 1.4.1.
- Pictures containing a single frontal face:
| Image size | Total pixels | Process time | FPS |
|---|---|---|---|
| 460x259 | 119,140 | 0.118 seconds | 8.5 |
| 561x561 | 314,721 | 0.227 seconds | 4.5 |
| 667x1000 | 667,000 | 0.456 seconds | 2.2 |
| 1920x1200 | 2,304,000 | 1.093 seconds | 0.9 |
| 4799x3599 | 17,271,601 | 8.798 seconds | 0.1 |
- Pictures containing 10 frontal faces:
| Image size | Total pixels | Process time | FPS |
|---|---|---|---|
| 474x224 | 106,176 | 0.185 seconds | 5.4 |
| 736x348 | 256,128 | 0.290 seconds | 3.4 |
| 2100x994 | 2,087,400 | 1.286 seconds | 0.7 |
By default the MTCNN bundles a face detection weights model.
The model is adapted from the Facenet's MTCNN implementation, merged in a single file located inside the folder 'data' relative to the module's path. It can be overriden by injecting it into the MTCNN() constructor during instantiation.
The model must be numpy-based containing the 3 main keys "pnet", "rnet" and "onet", having each of them the weights of each of the layers of the network.
For more reference about the network definition, take a close look at the paper from Zhang et al. (2016) [ZHANG2016].
| [ZHANG2016] | (1, 2) Zhang, K., Zhang, Z., Li, Z., and Qiao, Y. (2016). Joint face detection and alignment using multitask cascaded convolutional networks. IEEE Signal Processing Letters, 23(10):1499–1503. |