Repository for "A Convolutional Neural Network Cascade for Face Detection", implemented with Python interface.

This repo implemented the paper in python/tensorflow , providing the interface to contruct the cascade structure , including function of detection networks , calibration networks , image pyramids and non maximum supression .

• Thensorflow : Tensorflow installation guild
• Opencv : OpenCV installation guild

• Import Detector

# make sure detection.py is in program folder
from  detection import Detector

• Restore the pretrained model , download_link

# given paths from both models to Detector , and It will load the model on 
# your memory (or gpu memory).
det_mod_path = 'models/det_net_<epoch num>.ckpt'
cal_mod_path = 'models/cal_net_<epoch num>.ckpt'
detector = Detector(det_mod_path,cal_mod_path)

• Processing image pyramids

# bboxes is all bounding boxes of sliding windows , It’s include position
# and probability of face (default is -0.1)
# bboxes = [<xmin> , <ymin> , <xmax> , <ymax> , <probability>]
bboxes = detector.img_pyramids(image)

• Non Maximum Suppression

# iou_thresh is the overlapping threshold of iou in non maximum suppression 
# In returning bboxes , function will set the box’s probability = 0.0 which have 
# been filtered.
bboxes = detector.non_max_sup(bboxes,iou_thresh = 0.5)

• Predict the bounding boxes on detection/calibration net

# predict function will predict all the bounding boxes which probability is not 
# zero , function will set the box’s probability from prediction and return the 
# final bboxes .
# flags of net :  ‘net12’ , ‘net24’ , ‘net48’ , ‘net12_cal’ , ‘net24_cal’ , ‘net48_cal’ .
# threshold : the threshold of preditction.
bboxes = detector.predict(img,bboxes,net = 'net12',threshold = 0.9)

When I was training models , finding size of 12-net and 24-net was so hard to convergence , maybe the size of network is too small to learn pattern , so I change network size of net12 and net24 to 48*48 finally. But still confuse about it !

The accuracy of the calibration almost only 0.8 , It result the calibration of bounding box after network may making mistake , and bounding box will be removed in next stage , so sometimes I have better result without calibration net.

MIT LICENSE

Haoxiang Li, Zhe Lin, Xiaohui Shen, Jonathan Brandt, Gang Hua ; A Convolutional Neural Network Cascade for Face Detection ; The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015, pp. 5325-5334