How to install
$ pip install git+https://github.com/iitzco/faced.git
Soon to be available on
How to use
import cv2 from faced import FaceDetector from faced.utils import annotate_image face_detector = FaceDetector() img = cv2.imread(img_path) rgb_img = cv2.cvtColor(img.copy(), cv2.COLOR_BGR2RGB) # Receives RGB numpy image (HxWxC) and # returns (x_center, y_center, width, height, prob) tuples. bboxes = face_detector.predict(rgb_img, thresh) # Use this utils function to annotate the image. ann_img = annotate_image(img, bboxes) # Show the image cv2.imshow('image',ann_img) cv2.waitKey(0) cv2.destroyAllWindows()
As command-line program
# Detection on image saving the output $ faced --input imgs/demo.png --save
# Live webcam detection $ faced --input webcam
# Detection on video with low decision threshold $ faced --input imgs/demo.mp4 --threshold 0.5
faced --help for more information.
|CPU (i5 2015 MBP)||GPU (Nvidia TitanXP)|
|~5 FPS||> 70 FPS|
Comparison with Haar Cascades
Haar Cascades are one of the most used face detections models. Here's a comparison with OpenCV's implementation showing faced robustness.
faced is an ensemble of 2 deep neural networks (implemented using tensorflow) designed to run at Real Time speed in CPUs.
A custom fully convolutional neural network (FCNN) implementation based on YOLO. Takes a 288x288 RGB image and outputs a 9x9 grid where each cell can predict bounding boxes and probability of one face.
A custom standard CNN (Convolutions + Fully Connected layers) is used to take a face-containing rectangle and predict the face bounding box. This is a fine-tunning step. (outputs of Stage 1 model is not so accurate by itself, this is a corrector step that takes the each bouding box predicted from the previous step to improve bounding box quality.)
Why not just perform transfer learning on trained YOLO (or MobileNet+SSD) ?
Those models were designed to support multiclass detection (~80 classes). Because of this, these networks have to be powerfull enough to capture many different low and high level features that allow them to understand the patterns of very different classes. Powerful in this context means large amount of learnable parameters and hence big networks. These big networks cannot achieve real time performance on CPUs. 
This is an overkill for the simple task of just detecting faces. This work is a proof of concept that lighter networks can be designed to perform simpler tasks that do not require relatively large number of features.
 Those models cannot perform Real Time on CPU (YOLO at least). Even tiny-yolo version cannot achieve 1 fps on CPU (tested on 2015 MacBook Pro with 2.6 GHz Intel Core i5).
How was it trained?
Training was done with WIDER FACE dataset on Nvidia Titan XP GPU.
If you are interested in the training process and/or data preprocessing, just raise an
issueand we'll discuss it there.
How to run on GPU?
tensorflow-gpu instead of
This is not a Production ready system. Use it at your own risk.