Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Neural Networks | 邹进屹的博客

标签: | 发表时间:2018-12-20 13:37 | 作者:
出处:http://manutdzou.github.io

MTCNN

MTCNN算法由3个网络构成,分别是PNet,RNet以及ONet组成,其中PNet输出人脸位置和是人脸的概率,并且PNet是一个全卷积网络,在图像金字塔上不同尺度获得feature map每个pixel对应的人脸位置编码和人脸概率,然后通过阈值和NMS获得ROI人脸区域.第二个网络叫RNet主要对第一个网络获得的ROI区域进行refine,将第一个网络获得的所有ROI resize到24*24,重新分类获得所有ROI区域的人脸区域坐标和是人脸的概率。第三个网络叫ONet,对第二个CNN获得的人脸区域进行再次训练获得是否是人脸,人脸坐标以及五个特征点。

算法框架

1

2

测试图片

3

测试结果

4

以下项目时MTCNN的具体代码实现

项目地址:https://github.com/pangyupo/mxnet_mtcnn_face_detection

     # coding: utf-8

# main.py

import mxnet as mx
from mtcnn_detector import MtcnnDetector
import cv2
import os
import time

detector = MtcnnDetector(model_folder='model', ctx=mx.cpu(0), num_worker = 4 , accurate_landmark = False)


img = cv2.imread('test.jpg')

for i in range(4):
    t1 = time.time()
    results = detector.detect_face(img)
    print 'time: ',time.time() - t1

if results is not None:

    total_boxes = results[0]
    points = results[1]

    draw = img.copy()
    for b in total_boxes:
        cv2.rectangle(draw, (int(b[0]), int(b[1])), (int(b[2]), int(b[3])), (100, 255, 0))

    for p in points:
        for i in range(5):
            cv2.circle(draw, (p[i], p[i + 5]), 1, (0, 0, 255), 2)

    cv2.imshow("detection result", draw)
    cv2.waitKey(0)
    path = os.path.join('resutl'+'.jpg')
    cv2.imwrite(path,draw)

# --------------
# test on camera
# --------------
'''
camera = cv2.VideoCapture(0)
while True:
    grab, frame = camera.read()
    img = cv2.resize(frame, (320,180))

    t1 = time.time()
    results = detector.detect_face(img)
    print 'time: ',time.time() - t1

    if results is None:
        continue

    total_boxes = results[0]
    points = results[1]

    draw = img.copy()
    for b in total_boxes:
        cv2.rectangle(draw, (int(b[0]), int(b[1])), (int(b[2]), int(b[3])), (255, 255, 255))

    for p in points:
        for i in range(5):
            cv2.circle(draw, (p[i], p[i + 5]), 1, (255, 0, 0), 2)
    cv2.imshow("detection result", draw)
    cv2.waitKey(30)
'''
     # coding: utf-8
# mtcnn_detector.py
import os
import mxnet as mx
import numpy as np
import math
import cv2
from multiprocessing import Pool
from itertools import repeat
from itertools import izip
from helper import nms, adjust_input, generate_bbox, detect_first_stage_warpper

class MtcnnDetector(object):
    """
        Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Neural Networks
        see https://github.com/kpzhang93/MTCNN_face_detection_alignment
        this is a mxnet version
    """
    def __init__(self,
                 model_folder='.',
                 minsize = 20,
                 threshold = [0.6, 0.7, 0.7],
                 factor = 0.709,
                 num_worker = 1,
                 accurate_landmark = False,
                 ctx=mx.gpu()):
        """
            Initialize the detector

            Parameters:
            ----------
                model_folder : string
                    path for the models
                minsize : float number
                    minimal face to detect
                threshold : float number
                    detect threshold for 3 stages
                factor: float number
                    scale factor for image pyramid
                num_worker: int number
                    number of processes we use for first stage
                accurate_landmark: bool
                    use accurate landmark localization or not

        """
        self.num_worker = num_worker
        self.accurate_landmark = accurate_landmark

        # load 4 models from folder
        models = ['det1', 'det2', 'det3','det4']
        models = [ os.path.join(model_folder, f) for f in models]
        
        self.PNets = []
        for i in range(num_worker):
            workner_net = mx.model.FeedForward.load(models[0], 1, ctx=ctx)
            self.PNets.append(workner_net)

        self.Pool = Pool(num_worker)

        self.RNet = mx.model.FeedForward.load(models[1], 1, ctx=ctx)
        self.ONet = mx.model.FeedForward.load(models[2], 1, ctx=ctx)
        self.LNet = mx.model.FeedForward.load(models[3], 1, ctx=ctx)

        self.minsize   = float(minsize)
        self.factor    = float(factor)
        self.threshold = threshold


    def convert_to_square(self, bbox):
        """
            convert bbox to square

        Parameters:
        ----------
            bbox: numpy array , shape n x 5
                input bbox

        Returns:
        -------
            square bbox
        """
        square_bbox = bbox.copy()

        h = bbox[:, 3] - bbox[:, 1] + 1
        w = bbox[:, 2] - bbox[:, 0] + 1
        max_side = np.maximum(h,w)
        square_bbox[:, 0] = bbox[:, 0] + w*0.5 - max_side*0.5
        square_bbox[:, 1] = bbox[:, 1] + h*0.5 - max_side*0.5
        square_bbox[:, 2] = square_bbox[:, 0] + max_side - 1
        square_bbox[:, 3] = square_bbox[:, 1] + max_side - 1
        return square_bbox

    def calibrate_box(self, bbox, reg):
        """
            calibrate bboxes

        Parameters:
        ----------
            bbox: numpy array, shape n x 5
                input bboxes
            reg:  numpy array, shape n x 4
                bboxex adjustment

        Returns:
        -------
            bboxes after refinement

        """
        w = bbox[:, 2] - bbox[:, 0] + 1
        w = np.expand_dims(w, 1)
        h = bbox[:, 3] - bbox[:, 1] + 1
        h = np.expand_dims(h, 1)
        reg_m = np.hstack([w, h, w, h])
        aug = reg_m * reg
        bbox[:, 0:4] = bbox[:, 0:4] + aug
        return bbox

 
    def pad(self, bboxes, w, h):
        """
            pad the the bboxes, alse restrict the size of it

        Parameters:
        ----------
            bboxes: numpy array, n x 5
                input bboxes
            w: float number
                width of the input image
            h: float number
                height of the input image
        Returns :
        ------s
            dy, dx : numpy array, n x 1
                start point of the bbox in target image
            edy, edx : numpy array, n x 1
                end point of the bbox in target image
            y, x : numpy array, n x 1
                start point of the bbox in original image
            ex, ex : numpy array, n x 1
                end point of the bbox in original image
            tmph, tmpw: numpy array, n x 1
                height and width of the bbox

        """
        tmpw, tmph = bboxes[:, 2] - bboxes[:, 0] + 1,  bboxes[:, 3] - bboxes[:, 1] + 1
        num_box = bboxes.shape[0]

        dx , dy= np.zeros((num_box, )), np.zeros((num_box, ))
        edx, edy  = tmpw.copy()-1, tmph.copy()-1

        x, y, ex, ey = bboxes[:, 0], bboxes[:, 1], bboxes[:, 2], bboxes[:, 3]

        tmp_index = np.where(ex > w-1)
        edx[tmp_index] = tmpw[tmp_index] + w - 2 - ex[tmp_index]
        ex[tmp_index] = w - 1

        tmp_index = np.where(ey > h-1)
        edy[tmp_index] = tmph[tmp_index] + h - 2 - ey[tmp_index]
        ey[tmp_index] = h - 1

        tmp_index = np.where(x < 0)
        dx[tmp_index] = 0 - x[tmp_index]
        x[tmp_index] = 0

        tmp_index = np.where(y < 0)
        dy[tmp_index] = 0 - y[tmp_index]
        y[tmp_index] = 0

        return_list = [dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph]
        return_list = [item.astype(np.int32) for item in return_list]

        return  return_list

    def slice_index(self, number):
        """
            slice the index into (n,n,m), m < n
        Parameters:
        ----------
            number: int number
                number
        """
        def chunks(l, n):
            """Yield successive n-sized chunks from l."""
            for i in range(0, len(l), n):
                yield l[i:i + n]
        num_list = range(number)
        return list(chunks(num_list, self.num_worker))
        

    def detect_face(self, img):
        """
            detect face over img
        Parameters:
        ----------
            img: numpy array, bgr order of shape (1, 3, n, m)
                input image
        Retures:
        -------
            bboxes: numpy array, n x 5 (x1,y2,x2,y2,score)
                bboxes
            points: numpy array, n x 10 (x1, x2 ... x5, y1, y2 ..y5)
                landmarks
        """

        # check input
        MIN_DET_SIZE = 12

        if img is None:
            return None

        # only works for color image
        if len(img.shape) != 3:
            return None

        # detected boxes
        total_boxes = []

        height, width, _ = img.shape
        minl = min( height, width)

        # get all the valid scales
        scales = []
        m = MIN_DET_SIZE/self.minsize
        minl *= m
        factor_count = 0
        while minl > MIN_DET_SIZE:
            scales.append(m*self.factor**factor_count)
            minl *= self.factor
            factor_count += 1

        #############################################
        # first stage
        #############################################
        #for scale in scales:
        #    return_boxes = self.detect_first_stage(img, scale, 0)
        #    if return_boxes is not None:
        #        total_boxes.append(return_boxes)
        
        sliced_index = self.slice_index(len(scales))
        total_boxes = []
        for batch in sliced_index:
            local_boxes = self.Pool.map( detect_first_stage_warpper, \
                    izip(repeat(img), self.PNets[:len(batch)], [scales[i] for i in batch], repeat(self.threshold[0])) )
            total_boxes.extend(local_boxes)
        
        # remove the Nones 
        total_boxes = [ i for i in total_boxes if i is not None]

        if len(total_boxes) == 0:
            return None
        
        total_boxes = np.vstack(total_boxes)

        if total_boxes.size == 0:
            return None

        # merge the detection from first stage
        pick = nms(total_boxes[:, 0:5], 0.7, 'Union')
        total_boxes = total_boxes[pick]

        bbw = total_boxes[:, 2] - total_boxes[:, 0] + 1
        bbh = total_boxes[:, 3] - total_boxes[:, 1] + 1

        # refine the bboxes
        total_boxes = np.vstack([total_boxes[:, 0]+total_boxes[:, 5] * bbw,
                                 total_boxes[:, 1]+total_boxes[:, 6] * bbh,
                                 total_boxes[:, 2]+total_boxes[:, 7] * bbw,
                                 total_boxes[:, 3]+total_boxes[:, 8] * bbh,
                                 total_boxes[:, 4]
                                 ])

        total_boxes = total_boxes.T
        total_boxes = self.convert_to_square(total_boxes)
        total_boxes[:, 0:4] = np.round(total_boxes[:, 0:4])

        #############################################
        # second stage
        #############################################
        num_box = total_boxes.shape[0]

        # pad the bbox
        [dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph] = self.pad(total_boxes, width, height)
        # (3, 24, 24) is the input shape for RNet
        input_buf = np.zeros((num_box, 3, 24, 24), dtype=np.float32)

        for i in range(num_box):
            tmp = np.zeros((tmph[i], tmpw[i], 3), dtype=np.uint8)
            tmp[dy[i]:edy[i]+1, dx[i]:edx[i]+1, :] = img[y[i]:ey[i]+1, x[i]:ex[i]+1, :]
            input_buf[i, :, :, :] = adjust_input(cv2.resize(tmp, (24, 24)))

        output = self.RNet.predict(input_buf)

        # filter the total_boxes with threshold
        passed = np.where(output[1][:, 1] > self.threshold[1])
        total_boxes = total_boxes[passed]

        if total_boxes.size == 0:
            return None

        total_boxes[:, 4] = output[1][passed, 1].reshape((-1,))
        reg = output[0][passed]

        # nms
        pick = nms(total_boxes, 0.7, 'Union')
        total_boxes = total_boxes[pick]
        total_boxes = self.calibrate_box(total_boxes, reg[pick])
        total_boxes = self.convert_to_square(total_boxes)
        total_boxes[:, 0:4] = np.round(total_boxes[:, 0:4])

        #############################################
        # third stage
        #############################################
        num_box = total_boxes.shape[0]

        # pad the bbox
        [dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph] = self.pad(total_boxes, width, height)
        # (3, 48, 48) is the input shape for ONet
        input_buf = np.zeros((num_box, 3, 48, 48), dtype=np.float32)

        for i in range(num_box):
            tmp = np.zeros((tmph[i], tmpw[i], 3), dtype=np.float32)
            tmp[dy[i]:edy[i]+1, dx[i]:edx[i]+1, :] = img[y[i]:ey[i]+1, x[i]:ex[i]+1, :]
            input_buf[i, :, :, :] = adjust_input(cv2.resize(tmp, (48, 48)))

        output = self.ONet.predict(input_buf)

        # filter the total_boxes with threshold
        passed = np.where(output[2][:, 1] > self.threshold[2])
        total_boxes = total_boxes[passed]

        if total_boxes.size == 0:
            return None

        total_boxes[:, 4] = output[2][passed, 1].reshape((-1,))
        reg = output[1][passed]
        points = output[0][passed]

        # compute landmark points
        bbw = total_boxes[:, 2] - total_boxes[:, 0] + 1
        bbh = total_boxes[:, 3] - total_boxes[:, 1] + 1
        points[:, 0:5] = np.expand_dims(total_boxes[:, 0], 1) + np.expand_dims(bbw, 1) * points[:, 0:5]
        points[:, 5:10] = np.expand_dims(total_boxes[:, 1], 1) + np.expand_dims(bbh, 1) * points[:, 5:10]

        # nms
        total_boxes = self.calibrate_box(total_boxes, reg)
        pick = nms(total_boxes, 0.7, 'Min')
        total_boxes = total_boxes[pick]
        points = points[pick]
        
        if not self.accurate_landmark:
            return total_boxes, points

        #############################################
        # extended stage
        #############################################
        num_box = total_boxes.shape[0]
        patchw = np.maximum(total_boxes[:, 2]-total_boxes[:, 0]+1, total_boxes[:, 3]-total_boxes[:, 1]+1)
        patchw = np.round(patchw*0.25)

        # make it even
        patchw[np.where(np.mod(patchw,2) == 1)] += 1

        input_buf = np.zeros((num_box, 15, 24, 24), dtype=np.float32)
        for i in range(5):
            x, y = points[:, i], points[:, i+5]
            x, y = np.round(x-0.5*patchw), np.round(y-0.5*patchw)
            [dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph] = self.pad(np.vstack([x, y, x+patchw-1, y+patchw-1]).T,
                                                                    width,
                                                                    height)
            for j in range(num_box):
                tmpim = np.zeros((tmpw[j], tmpw[j], 3), dtype=np.float32)
                tmpim[dy[j]:edy[j]+1, dx[j]:edx[j]+1, :] = img[y[j]:ey[j]+1, x[j]:ex[j]+1, :]
                input_buf[j, i*3:i*3+3, :, :] = adjust_input(cv2.resize(tmpim, (24, 24)))

        output = self.LNet.predict(input_buf)

        pointx = np.zeros((num_box, 5))
        pointy = np.zeros((num_box, 5))

        for k in range(5):
            # do not make a large movement
            tmp_index = np.where(np.abs(output[k]-0.5) > 0.35)
            output[k][tmp_index[0]] = 0.5

            pointx[:, k] = np.round(points[:, k] - 0.5*patchw) + output[k][:, 0]*patchw
            pointy[:, k] = np.round(points[:, k+5] - 0.5*patchw) + output[k][:, 1]*patchw

        points = np.hstack([pointx, pointy])
        points = points.astype(np.int32)

        return total_boxes, points
     # coding: utf-8
# helper.py
import math
import cv2
import numpy as np


def nms(boxes, overlap_threshold, mode='Union'):
    """
        non max suppression

    Parameters:
    ----------
        box: numpy array n x 5
            input bbox array
        overlap_threshold: float number
            threshold of overlap
        mode: float number
            how to compute overlap ratio, 'Union' or 'Min'
    Returns:
    -------
        index array of the selected bbox
    """
    # if there are no boxes, return an empty list
    if len(boxes) == 0:
        return []

    # if the bounding boxes integers, convert them to floats
    if boxes.dtype.kind == "i":
        boxes = boxes.astype("float")

    # initialize the list of picked indexes
    pick = []

    # grab the coordinates of the bounding boxes
    x1, y1, x2, y2, score = [boxes[:, i] for i in range(5)]

    area = (x2 - x1 + 1) * (y2 - y1 + 1)
    idxs = np.argsort(score)

    # keep looping while some indexes still remain in the indexes list
    while len(idxs) > 0:
        # grab the last index in the indexes list and add the index value to the list of picked indexes
        last = len(idxs) - 1
        i = idxs[last]
        pick.append(i)

        xx1 = np.maximum(x1[i], x1[idxs[:last]])
        yy1 = np.maximum(y1[i], y1[idxs[:last]])
        xx2 = np.minimum(x2[i], x2[idxs[:last]])
        yy2 = np.minimum(y2[i], y2[idxs[:last]])

        # compute the width and height of the bounding box
        w = np.maximum(0, xx2 - xx1 + 1)
        h = np.maximum(0, yy2 - yy1 + 1)

        inter = w * h
        if mode == 'Min':
            overlap = inter / np.minimum(area[i], area[idxs[:last]])
        else:
            overlap = inter / (area[i] + area[idxs[:last]] - inter)

        # delete all indexes from the index list that have
        idxs = np.delete(idxs, np.concatenate(([last],
                                               np.where(overlap > overlap_threshold)[0])))

    return pick

def adjust_input(in_data):
    """
        adjust the input from (h, w, c) to ( 1, c, h, w) for network input

    Parameters:
    ----------
        in_data: numpy array of shape (h, w, c)
            input data
    Returns:
    -------
        out_data: numpy array of shape (1, c, h, w)
            reshaped array
    """
    if in_data.dtype is not np.dtype('float32'):
        out_data = in_data.astype(np.float32)
    else:
        out_data = in_data

    out_data = out_data.transpose((2,0,1))
    out_data = np.expand_dims(out_data, 0)
    out_data = (out_data - 127.5)*0.0078125
    return out_data

def generate_bbox(map, reg, scale, threshold):
     """
         generate bbox from feature map
     Parameters:
     ----------
         map: numpy array , n x m x 1
             detect score for each position
         reg: numpy array , n x m x 4
             bbox
         scale: float number
             scale of this detection
         threshold: float number
             detect threshold
     Returns:
     -------
         bbox array
     """
     stride = 2
     cellsize = 12

     t_index = np.where(map>threshold)

     # find nothing
     if t_index[0].size == 0:
         return np.array([])

     dx1, dy1, dx2, dy2 = [reg[0, i, t_index[0], t_index[1]] for i in range(4)]

     reg = np.array([dx1, dy1, dx2, dy2])
     score = map[t_index[0], t_index[1]]
     boundingbox = np.vstack([np.round((stride*t_index[1]+1)/scale),
                              np.round((stride*t_index[0]+1)/scale),
                              np.round((stride*t_index[1]+1+cellsize)/scale),
                              np.round((stride*t_index[0]+1+cellsize)/scale),
                              score,
                              reg])

     return boundingbox.T


def detect_first_stage(img, net, scale, threshold):
    """
        run PNet for first stage
    
    Parameters:
    ----------
        img: numpy array, bgr order
            input image
        scale: float number
            how much should the input image scale
        net: PNet
            worker
    Returns:
    -------
        total_boxes : bboxes
    """
    height, width, _ = img.shape
    hs = int(math.ceil(height * scale))
    ws = int(math.ceil(width * scale))
    
    im_data = cv2.resize(img, (ws,hs))
    
    # adjust for the network input
    input_buf = adjust_input(im_data)
    output = net.predict(input_buf)
    boxes = generate_bbox(output[1][0,1,:,:], output[0], scale, threshold)

    if boxes.size == 0:
        return None

    # nms
    pick = nms(boxes[:,0:5], 0.5, mode='Union')
    boxes = boxes[pick]
    return boxes

def detect_first_stage_warpper( args ):
    return detect_first_stage(*args)

caffe的实现版本

_init_paths.py

     import os.path as osp
import sys

def add_path(path):
    if path not in sys.path:
        sys.path.insert(0, path)

caffe_path = '/home/zou/caffe'

# Add caffe to PYTHONPATH
caffe_path = osp.join(caffe_path, 'python')
add_path(caffe_path)

demo.py

     #!/usr/bin/env python
# -*- coding: utf-8 -*-

import _init_paths
import caffe
import cv2
import numpy as np

def bbreg(boundingbox, reg):
    reg = reg.T 
    
    # calibrate bouding boxes
    if reg.shape[1] == 1:
        pass # reshape of reg
    w = boundingbox[:,2] - boundingbox[:,0] + 1
    h = boundingbox[:,3] - boundingbox[:,1] + 1

    bb0 = boundingbox[:,0] + reg[:,0]*w
    bb1 = boundingbox[:,1] + reg[:,1]*h
    bb2 = boundingbox[:,2] + reg[:,2]*w
    bb3 = boundingbox[:,3] + reg[:,3]*h
    
    boundingbox[:,0:4] = np.array([bb0, bb1, bb2, bb3]).T
    #print "bb", boundingbox
    return boundingbox


def pad(boxesA, w, h):
    boxes = boxesA.copy() # shit, value parameter!!!
    
    tmph = boxes[:,3] - boxes[:,1] + 1
    tmpw = boxes[:,2] - boxes[:,0] + 1
    numbox = boxes.shape[0]

    dx = np.ones(numbox)
    dy = np.ones(numbox)
    edx = tmpw 
    edy = tmph

    x = boxes[:,0:1][:,0]
    y = boxes[:,1:2][:,0]
    ex = boxes[:,2:3][:,0]
    ey = boxes[:,3:4][:,0]
   
   
    tmp = np.where(ex > w)[0]
    if tmp.shape[0] != 0:
        edx[tmp] = -ex[tmp] + w-1 + tmpw[tmp]
        ex[tmp] = w-1

    tmp = np.where(ey > h)[0]
    if tmp.shape[0] != 0:
        edy[tmp] = -ey[tmp] + h-1 + tmph[tmp]
        ey[tmp] = h-1

    tmp = np.where(x < 1)[0]
    if tmp.shape[0] != 0:
        dx[tmp] = 2 - x[tmp]
        x[tmp] = np.ones_like(x[tmp])

    tmp = np.where(y < 1)[0]
    if tmp.shape[0] != 0:
        dy[tmp] = 2 - y[tmp]
        y[tmp] = np.ones_like(y[tmp])
    
    # for python index from 0, while matlab from 1
    dy = np.maximum(0, dy-1)
    dx = np.maximum(0, dx-1)
    y = np.maximum(0, y-1)
    x = np.maximum(0, x-1)
    edy = np.maximum(0, edy-1)
    edx = np.maximum(0, edx-1)
    ey = np.maximum(0, ey-1)
    ex = np.maximum(0, ex-1)

    #print 'boxes', boxes
    return [dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph]



def rerec(bboxA):
    # convert bboxA to square
    w = bboxA[:,2] - bboxA[:,0]
    h = bboxA[:,3] - bboxA[:,1]
    l = np.maximum(w,h).T

    bboxA[:,0] = bboxA[:,0] + w*0.5 - l*0.5
    bboxA[:,1] = bboxA[:,1] + h*0.5 - l*0.5 
    bboxA[:,2:4] = bboxA[:,0:2] + np.repeat([l], 2, axis = 0).T 
    return bboxA


def nms(boxes, threshold, type):
    """nms
    :boxes: [:,0:5]
    :threshold: 0.5 like
    :type: 'Min' or others
    :returns: TODO
    """
    if boxes.shape[0] == 0:
        return np.array([])
    x1 = boxes[:,0]
    y1 = boxes[:,1]
    x2 = boxes[:,2]
    y2 = boxes[:,3]
    s = boxes[:,4]
    area = np.multiply(x2-x1+1, y2-y1+1)
    I = np.array(s.argsort()) # read s using I
    
    pick = [];
    while len(I) > 0:
        xx1 = np.maximum(x1[I[-1]], x1[I[0:-1]])
        yy1 = np.maximum(y1[I[-1]], y1[I[0:-1]])
        xx2 = np.minimum(x2[I[-1]], x2[I[0:-1]])
        yy2 = np.minimum(y2[I[-1]], y2[I[0:-1]])
        w = np.maximum(0.0, xx2 - xx1 + 1)
        h = np.maximum(0.0, yy2 - yy1 + 1)
        inter = w * h
        if type == 'Min':
            o = inter / np.minimum(area[I[-1]], area[I[0:-1]])
        else:
            o = inter / (area[I[-1]] + area[I[0:-1]] - inter)
        pick.append(I[-1])
        I = I[np.where( o <= threshold)[0]]
    return pick


def generateBoundingBox(map, reg, scale, t):
    stride = 2
    cellsize = 12
    map = map.T
    dx1 = reg[0,:,:].T
    dy1 = reg[1,:,:].T
    dx2 = reg[2,:,:].T
    dy2 = reg[3,:,:].T
    (x, y) = np.where(map >= t)

    yy = y
    xx = x
    
    '''
    if y.shape[0] == 1: # only one point exceed threshold
        y = y.T
        x = x.T
        score = map[x,y].T
        dx1 = dx1.T
        dy1 = dy1.T
        dx2 = dx2.T
        dy2 = dy2.T
        # a little stange, when there is only one bb created by PNet
        
        #print "1: x,y", x,y
        a = (x*map.shape[1]) + (y+1)
        x = a/map.shape[0]
        y = a%map.shape[0] - 1
        #print "2: x,y", x,y
    else:
        score = map[x,y]
    '''

    score = map[x,y]
    reg = np.array([dx1[x,y], dy1[x,y], dx2[x,y], dy2[x,y]])

    if reg.shape[0] == 0:
        pass
    boundingbox = np.array([yy, xx]).T

    bb1 = np.fix((stride * (boundingbox) + 1) / scale).T # matlab index from 1, so with "boundingbox-1"
    bb2 = np.fix((stride * (boundingbox) + cellsize - 1 + 1) / scale).T # while python don't have to
    score = np.array([score])

    boundingbox_out = np.concatenate((bb1, bb2, score, reg), axis=0)

    return boundingbox_out.T



def drawBoxes(im, boxes):
    x1 = boxes[:,0]
    y1 = boxes[:,1]
    x2 = boxes[:,2]
    y2 = boxes[:,3]
    for i in range(x1.shape[0]):
        cv2.rectangle(im, (int(x1[i]), int(y1[i])), (int(x2[i]), int(y2[i])), (0,255,0), 1)
    return im

def drawPoints(im,points):
    for i in range(points.shape[0]):
        left_eye = (int(points[i][0]),int(points[i][5]))
        cv2.circle(im, left_eye,2, (0,0,255), 2)

        right_eye = (int(points[i][1]),int(points[i][6]))
        cv2.circle(im, right_eye,2, (0,0,255), 2)

        nose = (int(points[i][2]),int(points[i][7]))
        cv2.circle(im, nose,2, (0,0,255), 2)

        left_mouth = (int(points[i][3]),int(points[i][8]))
        cv2.circle(im, left_mouth,2, (0,0,255), 2)

        right_mouth = (int(points[i][4]),int(points[i][9]))
        cv2.circle(im, right_mouth,2, (0,0,255), 2)
    return im

def drawPatch(im,points):
    for i in range(points.shape[0]):
        left_eye = (int(points[i][0]),int(points[i][5]))

        right_eye = (int(points[i][1]),int(points[i][6]))

        nose = (int(points[i][2]),int(points[i][7]))

        left_mouth = (int(points[i][3]),int(points[i][8]))

        right_mouth = (int(points[i][4]),int(points[i][9]))

        eye_length = np.sqrt((left_eye[0]-right_eye[0])*(left_eye[0]-right_eye[0])+(left_eye[1]-right_eye[1])*(left_eye[1]-right_eye[1]))
	mouth_length = np.sqrt((left_mouth[0]-right_mouth[0])*(left_mouth[0]-right_mouth[0])+(left_mouth[1]-right_mouth[1])*(left_mouth[1]-right_mouth[1]))

	t11_x = left_eye[0]
        t11_y = left_eye[1] - 0.8*eye_length
        t12_x = left_eye[0] + eye_length
        t12_y = left_eye[1] - 0.4*eye_length
        cv2.rectangle(im, (int(t11_x), int(t11_y)), (int(t12_x), int(t12_y)), (0,255,0), 1)

	t21_x = (left_eye[0] + right_eye[0])/2 - 0.15*eye_length
        t21_y = (left_eye[1] + right_eye[1])/2 - 0.3*eye_length
        t22_x = (left_eye[0] + right_eye[0])/2 + 0.15*eye_length
        t22_y = (left_eye[1] + right_eye[1])/2
	cv2.rectangle(im, (int(t21_x), int(t21_y)), (int(t22_x), int(t22_y)), (0,255,0), 1)

	t31_x = (left_eye[0] +nose[0])/2-0.1*eye_length
        t31_y = ((left_eye[1] + right_eye[1])/2 + nose[1])/2 - 0.1*eye_length
        t32_x = (left_eye[0] +nose[0])/2+0.1*eye_length
        t32_y = ((left_eye[1] + right_eye[1])/2 + nose[1])/2 + 0.1*eye_length
	cv2.rectangle(im, (int(t31_x), int(t31_y)), (int(t32_x), int(t32_y)), (0,255,0), 1)

	t41_x = (right_eye[0] +nose[0])/2-0.1*eye_length
        t41_y = ((left_eye[1] + right_eye[1])/2 + nose[1])/2 - 0.1*eye_length
        t42_x = (right_eye[0] +nose[0])/2+0.1*eye_length
        t42_y = ((left_eye[1] + right_eye[1])/2 + nose[1])/2 + 0.1*eye_length
	cv2.rectangle(im, (int(t41_x), int(t41_y)), (int(t42_x), int(t42_y)), (0,255,0), 1)

	t51_x = nose[0]-0.1*eye_length
        t51_y = nose[1]-0.1*eye_length
        t52_x = nose[0]+0.1*eye_length
        t52_y = nose[1]+0.1*eye_length
	cv2.rectangle(im, (int(t51_x), int(t51_y)), (int(t52_x), int(t52_y)), (0,255,0), 1)

	t61_x = nose[0]-0.1*eye_length
        t61_y = nose[1]+0.2*eye_length
        t62_x = nose[0]+0.1*eye_length
        t62_y = nose[1]+0.4*eye_length
	cv2.rectangle(im, (int(t61_x), int(t61_y)), (int(t62_x), int(t62_y)), (0,255,0), 1)

	t71_x = left_mouth[0] - 0.7*mouth_length
        t71_y = left_mouth[1] - mouth_length
        t72_x = left_mouth[0]
        t72_y = left_mouth[1]
	cv2.rectangle(im, (int(t71_x), int(t71_y)), (int(t72_x), int(t72_y)), (0,255,0), 1)

	t81_x = right_mouth[0]
        t81_y = right_mouth[1] - mouth_length
        t82_x = right_mouth[0] + 0.7*mouth_length
        t82_y = left_mouth[1]
	cv2.rectangle(im, (int(t81_x), int(t81_y)), (int(t82_x), int(t82_y)), (0,255,0), 1)

    return im


from time import time
_tstart_stack = []
def tic():
    _tstart_stack.append(time())
def toc(fmt="Elapsed: %s s"):
    print fmt % (time()-_tstart_stack.pop())


def detect_face(img, minsize, PNet, RNet, ONet, threshold, fastresize, factor):
    
    img2 = img.copy()

    factor_count = 0
    total_boxes = np.zeros((0,9), np.float)
    points = []
    h = img.shape[0]
    w = img.shape[1]
    minl = min(h, w)
    img = img.astype(float)
    m = 12.0/minsize
    minl = minl*m

    # create scale pyramid
    scales = []
    while minl >= 12:
        scales.append(m * pow(factor, factor_count))
        minl *= factor
        factor_count += 1
    
    # first stage
    for scale in scales:
        hs = int(np.ceil(h*scale))
        ws = int(np.ceil(w*scale))

        if fastresize:
            im_data = (img-127.5)*0.0078125 # [0,255] -> [-1,1]
            im_data = cv2.resize(im_data, (ws,hs)) # default is bilinear
        else: 
            im_data = cv2.resize(img, (ws,hs)) # default is bilinear
            im_data = (im_data-127.5)*0.0078125 # [0,255] -> [-1,1]

        im_data = np.swapaxes(im_data, 0, 2)
        im_data = np.array([im_data], dtype = np.float)
        PNet.blobs['data'].reshape(1, 3, ws, hs)
        PNet.blobs['data'].data[...] = im_data
        out = PNet.forward()
    
        boxes = generateBoundingBox(out['prob1'][0,1,:,:], out['conv4-2'][0], scale, threshold[0])
        if boxes.shape[0] != 0:
            pick = nms(boxes, 0.5, 'Union')
            if len(pick) > 0 :
                boxes = boxes[pick, :]

        if boxes.shape[0] != 0:
            total_boxes = np.concatenate((total_boxes, boxes), axis=0)

    #####
    # 1 #
    #####
    numbox = total_boxes.shape[0]
    if numbox > 0:
        # nms
        pick = nms(total_boxes, 0.7, 'Union')
        total_boxes = total_boxes[pick, :]
        
        # revise and convert to square
        regh = total_boxes[:,3] - total_boxes[:,1]
        regw = total_boxes[:,2] - total_boxes[:,0]
        t1 = total_boxes[:,0] + total_boxes[:,5]*regw
        t2 = total_boxes[:,1] + total_boxes[:,6]*regh
        t3 = total_boxes[:,2] + total_boxes[:,7]*regw
        t4 = total_boxes[:,3] + total_boxes[:,8]*regh
        t5 = total_boxes[:,4]
        total_boxes = np.array([t1,t2,t3,t4,t5]).T

        total_boxes = rerec(total_boxes) # convert box to square
        total_boxes[:,0:4] = np.fix(total_boxes[:,0:4])
        [dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph] = pad(total_boxes, w, h)

    #print total_boxes.shape
    #print total_boxes

    numbox = total_boxes.shape[0]
    if numbox > 0:
        # second stage
        # construct input for RNet
        tempimg = np.zeros((numbox, 24, 24, 3)) # (24, 24, 3, numbox)
        for k in range(numbox):
            tmp = np.zeros((tmph[k], tmpw[k],3))
            tmp[dy[k]:edy[k]+1, dx[k]:edx[k]+1] = img[y[k]:ey[k]+1, x[k]:ex[k]+1]
            tempimg[k,:,:,:] = cv2.resize(tmp, (24, 24))

        tempimg = (tempimg-127.5)*0.0078125 # done in imResample function wrapped by python

        # RNet

        tempimg = np.swapaxes(tempimg, 1, 3)
        
        RNet.blobs['data'].reshape(numbox, 3, 24, 24)
        RNet.blobs['data'].data[...] = tempimg
        out = RNet.forward()

        score = out['prob1'][:,1]
        pass_t = np.where(score>threshold[1])[0]
        
        score =  np.array([score[pass_t]]).T
        total_boxes = np.concatenate( (total_boxes[pass_t, 0:4], score), axis = 1)
        
        mv = out['conv5-2'][pass_t, :].T
        if total_boxes.shape[0] > 0:
            pick = nms(total_boxes, 0.7, 'Union')
            if len(pick) > 0 :
                total_boxes = total_boxes[pick, :]
                total_boxes = bbreg(total_boxes, mv[:, pick])
                total_boxes = rerec(total_boxes)
            
        #####
        # 2 #
        #####
        numbox = total_boxes.shape[0]
        if numbox > 0:
            # third stage
            
            total_boxes = np.fix(total_boxes)
            [dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph] = pad(total_boxes, w, h)

            tempimg = np.zeros((numbox, 48, 48, 3))
            for k in range(numbox):
                tmp = np.zeros((tmph[k], tmpw[k],3))
                tmp[dy[k]:edy[k]+1, dx[k]:edx[k]+1] = img[y[k]:ey[k]+1, x[k]:ex[k]+1]
                tempimg[k,:,:,:] = cv2.resize(tmp, (48, 48))
            tempimg = (tempimg-127.5)*0.0078125 # [0,255] -> [-1,1]
                
            # ONet
            tempimg = np.swapaxes(tempimg, 1, 3)
            ONet.blobs['data'].reshape(numbox, 3, 48, 48)
            ONet.blobs['data'].data[...] = tempimg
            out = ONet.forward()
            
            score = out['prob1'][:,1]
            points = out['conv6-3']
            pass_t = np.where(score>threshold[2])[0]
            points = points[pass_t, :]
            score = np.array([score[pass_t]]).T
            total_boxes = np.concatenate( (total_boxes[pass_t, 0:4], score), axis=1)
            
            mv = out['conv6-2'][pass_t, :].T
            w = total_boxes[:,3] - total_boxes[:,1] + 1
            h = total_boxes[:,2] - total_boxes[:,0] + 1

            points[:, 0:5] = np.tile(w, (5,1)).T * points[:, 0:5] + np.tile(total_boxes[:,0], (5,1)).T - 1 
            points[:, 5:10] = np.tile(h, (5,1)).T * points[:, 5:10] + np.tile(total_boxes[:,1], (5,1)).T -1

            if total_boxes.shape[0] > 0:
                total_boxes = bbreg(total_boxes, mv[:,:])
                pick = nms(total_boxes, 0.7, 'Min')
                if len(pick) > 0 :
                    total_boxes = total_boxes[pick, :]
                    points = points[pick, :]

    #####
    # 3 #
    #####
    return total_boxes, points




    
def initFaceDetector():
    minsize = 20
    caffe_model_path = "/home/duino/iactive/mtcnn/model"
    threshold = [0.6, 0.7, 0.7]
    factor = 0.709
    caffe.set_mode_cpu()
    PNet = caffe.Net(caffe_model_path+"/det1.prototxt", caffe_model_path+"/det1.caffemodel", caffe.TEST)
    RNet = caffe.Net(caffe_model_path+"/det2.prototxt", caffe_model_path+"/det2.caffemodel", caffe.TEST)
    ONet = caffe.Net(caffe_model_path+"/det3.prototxt", caffe_model_path+"/det3.caffemodel", caffe.TEST)
    return (minsize, PNet, RNet, ONet, threshold, factor)

def haveFace(img, facedetector):
    minsize = facedetector[0]
    PNet = facedetector[1]
    RNet = facedetector[2]
    ONet = facedetector[3]
    threshold = facedetector[4]
    factor = facedetector[5]
    
    if max(img.shape[0], img.shape[1]) < minsize:
        return False, []

    img_matlab = img.copy()
    tmp = img_matlab[:,:,2].copy()
    img_matlab[:,:,2] = img_matlab[:,:,0]
    img_matlab[:,:,0] = tmp
    
    #tic()
    boundingboxes, points = detect_face(img_matlab, minsize, PNet, RNet, ONet, threshold, False, factor)
    #toc()
    containFace = (True, False)[boundingboxes.shape[0]==0]
    return containFace, boundingboxes

def main():
    minsize = 50

    caffe_model_path = "/home/zou/mtcnn/model"

    threshold = [0.6, 0.7, 0.7]
    factor = 0.709
    
    caffe.set_mode_gpu()
    PNet = caffe.Net(caffe_model_path+"/det1.prototxt", caffe_model_path+"/det1.caffemodel", caffe.TEST)
    RNet = caffe.Net(caffe_model_path+"/det2.prototxt", caffe_model_path+"/det2.caffemodel", caffe.TEST)
    ONet = caffe.Net(caffe_model_path+"/det3.prototxt", caffe_model_path+"/det3.caffemodel", caffe.TEST)

    camera = cv2.VideoCapture(0)
    while True:
        _, img = camera.read()
        h,w = img.shape[:2]
        if h>=w:
            w = int(w/(h/500.0))
            h = 500;
        else:
            h = int(h/(w/500.0))
            w = 500

        img = cv2.resize(img,(w,h))
        img_matlab = img.copy()
        tmp = img_matlab[:,:,2].copy()
        img_matlab[:,:,2] = img_matlab[:,:,0]
        img_matlab[:,:,0] = tmp

        # check rgb position
        tic()
        boundingboxes, points = detect_face(img_matlab, minsize, PNet, RNet, ONet, threshold, False, factor)
        toc()

        if (len(boundingboxes)>0)&1:
            img = drawBoxes(img, boundingboxes)
            img = drawPoints(img, points)
        if (len(boundingboxes)>0)&0:
            img = drawBoxes(img, boundingboxes)
            img = drawPatch(img, points)
        cv2.imshow('img', img)

        ch = cv2.waitKey(1)
        if ch == 27:
            break


        #if boundingboxes.shape[0] > 0:
        #    error.append[imgpath]
    #print error

if __name__ == "__main__":
    main()

C++实现

stdafx.h

   // stdafx.h : 标准系统包含文件的包含文件,
// 或是经常使用但不常更改的
// 特定于项目的包含文件
//

#pragma once

#include "targetver.h"

#include <stdio.h>
#include <tchar.h>

#include <caffe/caffe.hpp>
#ifdef USE_OPENCV
#include <opencv2/core/core.hpp>
#include <opencv2/highgui/highgui.hpp>
#include <opencv2/imgproc/imgproc.hpp>
#include <opencv2/videoio.hpp>
#endif  // USE_OPENCV
#include <algorithm>
#include <iosfwd>
#include <memory>
#include <string>
#include <utility>
#include <vector>

#ifdef WITH_PYTHON_LAYER
#include <boost/python.hpp>
#endif
#include <string>
#include <vector>

#include "caffe/layer.hpp"
#include "caffe/layer_factory.hpp"
#include "caffe/layers/input_layer.hpp"
#include "caffe/layers/inner_product_layer.hpp"
#include "caffe/layers/prelu_layer.hpp"
#include "caffe/layers/conv_layer.hpp"
#include "caffe/layers/pooling_layer.hpp"
#include "caffe/layers/softmax_layer.hpp"
#include "caffe/layers/memory_data_layer.hpp"
#include "caffe/layers/dropout_layer.hpp"
#include "caffe/proto/caffe.pb.h"

#ifdef USE_CUDNN
#include "caffe/layers/cudnn_conv_layer.hpp"
#include "caffe/layers/cudnn_pooling_layer.hpp"
#include "caffe/layers/cudnn_relu_layer.hpp"
#include "caffe/layers/cudnn_softmax_layer.hpp"
#endif

#ifdef WITH_PYTHON_LAYER
#include "caffe/layers/python_layer.hpp"
#endif

using namespace caffe;  // NOLINT(build/namespaces)

extern INSTANTIATE_CLASS(InputLayer);
extern INSTANTIATE_CLASS(InnerProductLayer);
extern INSTANTIATE_CLASS(ConvolutionLayer);
extern INSTANTIATE_CLASS(PReLULayer);
extern INSTANTIATE_CLASS(PoolingLayer);
extern INSTANTIATE_CLASS(SoftmaxLayer);
extern INSTANTIATE_CLASS(MemoryDataLayer);
extern INSTANTIATE_CLASS(DropoutLayer);
// TODO:  在此处引用程序需要的其他头文件

MTCNN.cpp

   // MTCNN_VS2015.cpp : 定义控制台应用程序的入口点。
//

#include "stdafx.h"
// c++
#include <string>
#include <vector>
// boost
#include "boost/make_shared.hpp"
//#define CPU_ONLY
using namespace caffe;

#define FROM_VIDEO 1

string resultdir = "result";

typedef struct FaceRect {
	float x1;
	float y1;
	float x2;
	float y2;
	float score; /**< Larger score should mean higher confidence. */
} FaceRect;

typedef struct FacePts {
	float x[5], y[5];
} FacePts;

typedef struct FaceInfo {
	FaceRect bbox;
	cv::Vec4f regression;
	FacePts facePts;
	double roll;
	double pitch;
	double yaw;
} FaceInfo;

template<typename Dtype>
Dtype max(Dtype x, Dtype y)
{
	return x>=y ? x : y;
}

template<typename Dtype>
Dtype min(Dtype x, Dtype y)
{
	return x < y ? x : y;
}

class MTCNN {
public:
	MTCNN(const string& proto_model_dir);
	void Detect(const cv::Mat& img, std::vector<FaceInfo> &faceInfo, int minSize, double* threshold, double factor);

private:
	bool CvMatToDatumSignalChannel(const cv::Mat& cv_mat, Datum* datum);
	//void Preprocess(const cv::Mat& img,
	//	std::vector<cv::Mat>* input_channels);
	void WrapInputLayer(std::vector<cv::Mat>* input_channels, Blob<float>* input_layer,
		const int height, const int width);
	//void SetMean();
	void GenerateBoundingBox(Blob<float>* confidence, Blob<float>* reg,
		float scale, float thresh, int image_width, int image_height);
	void ClassifyFace_MulImage(const std::vector<FaceInfo> &regressed_rects, cv::Mat &sample_single,
		boost::shared_ptr<Net<float> >& net, double thresh, char netName);
	std::vector<FaceInfo> NonMaximumSuppression(std::vector<FaceInfo>& bboxes, float thresh, char methodType);
	void Bbox2Square(std::vector<FaceInfo>& bboxes);
	void Padding(int img_w, int img_h);
	std::vector<FaceInfo> BoxRegress(std::vector<FaceInfo> &faceInfo_, int stage);
	//void RegressPoint(const std::vector<FaceInfo>& faceInfo);

private:
	boost::shared_ptr<Net<float> > PNet_;
	boost::shared_ptr<Net<float> > RNet_;
	boost::shared_ptr<Net<float> > ONet_;

	// x1,y1,x2,t2 and score
	std::vector<FaceInfo> condidate_rects_;
	std::vector<FaceInfo> total_boxes_;
	std::vector<FaceInfo> regressed_rects_;
	std::vector<FaceInfo> regressed_pading_;

	std::vector<cv::Mat> crop_img_;
	int curr_feature_map_w_;
	int curr_feature_map_h_;
	int num_channels_;
};

// compare score
bool CompareBBox(const FaceInfo & a, const FaceInfo & b) {
	return a.bbox.score > b.bbox.score;
}

// methodType : u is IoU(Intersection Over Union)
// methodType : m is IoM(Intersection Over Maximum)
std::vector<FaceInfo> MTCNN::NonMaximumSuppression(std::vector<FaceInfo>& bboxes,
	float thresh, char methodType) {
	std::vector<FaceInfo> bboxes_nms;
	std::sort(bboxes.begin(), bboxes.end(), CompareBBox);

	int32_t select_idx = 0;
	int32_t num_bbox = static_cast<int32_t>(bboxes.size());
	std::vector<int32_t> mask_merged(num_bbox, 0);
	bool all_merged = false;

	while (!all_merged) {
		while (select_idx < num_bbox && mask_merged[select_idx] == 1)
			select_idx++;
		if (select_idx == num_bbox) {
			all_merged = true;
			continue;
		}

		bboxes_nms.push_back(bboxes[select_idx]);
		mask_merged[select_idx] = 1;

		FaceRect select_bbox = bboxes[select_idx].bbox;
		float area1 = static_cast<float>((select_bbox.x2 - select_bbox.x1 + 1) * (select_bbox.y2 - select_bbox.y1 + 1));
		float x1 = static_cast<float>(select_bbox.x1);
		float y1 = static_cast<float>(select_bbox.y1);
		float x2 = static_cast<float>(select_bbox.x2);
		float y2 = static_cast<float>(select_bbox.y2);

		select_idx++;
		for (int32_t i = select_idx; i < num_bbox; i++) {
			if (mask_merged[i] == 1)
				continue;

			FaceRect& bbox_i = bboxes[i].bbox;
			float x = std::max<float>(x1, static_cast<float>(bbox_i.x1));
			float y = std::max<float>(y1, static_cast<float>(bbox_i.y1));
			float w = std::min<float>(x2, static_cast<float>(bbox_i.x2)) - x + 1;
			float h = std::min<float>(y2, static_cast<float>(bbox_i.y2)) - y + 1;
			if (w <= 0 || h <= 0)
				continue;

			float area2 = static_cast<float>((bbox_i.x2 - bbox_i.x1 + 1) * (bbox_i.y2 - bbox_i.y1 + 1));
			float area_intersect = w * h;

			switch (methodType) {
			case 'u':
				if (static_cast<float>(area_intersect) / (area1 + area2 - area_intersect) > thresh)
					mask_merged[i] = 1;
				break;
			case 'm':
				if (static_cast<float>(area_intersect) / min(area1, area2) > thresh)
					mask_merged[i] = 1;
				break;
			default:
				break;
			}
		}
	}
	return bboxes_nms;
}

void MTCNN::Bbox2Square(std::vector<FaceInfo>& bboxes) {
	for (int i = 0; i < bboxes.size(); i++) {
		float h = bboxes[i].bbox.x2 - bboxes[i].bbox.x1;
		float w = bboxes[i].bbox.y2 - bboxes[i].bbox.y1;
		float side = h > w ? h : w;
		bboxes[i].bbox.x1 += (h - side)*0.5;
		bboxes[i].bbox.y1 += (w - side)*0.5;

		bboxes[i].bbox.x2 = (int)(bboxes[i].bbox.x1 + side);
		bboxes[i].bbox.y2 = (int)(bboxes[i].bbox.y1 + side);
		bboxes[i].bbox.x1 = (int)(bboxes[i].bbox.x1);
		bboxes[i].bbox.y1 = (int)(bboxes[i].bbox.y1);

	}
}

std::vector<FaceInfo> MTCNN::BoxRegress(std::vector<FaceInfo>& faceInfo, int stage) {
	std::vector<FaceInfo> bboxes;
	for (int bboxId = 0; bboxId < faceInfo.size(); bboxId++) {
		FaceRect faceRect;
		FaceInfo tempFaceInfo;
		float regw = faceInfo[bboxId].bbox.y2 - faceInfo[bboxId].bbox.y1;
		regw += (stage == 1) ? 0 : 1;
		float regh = faceInfo[bboxId].bbox.x2 - faceInfo[bboxId].bbox.x1;
		regh += (stage == 1) ? 0 : 1;
		faceRect.y1 = faceInfo[bboxId].bbox.y1 + regw * faceInfo[bboxId].regression[0];
		faceRect.x1 = faceInfo[bboxId].bbox.x1 + regh * faceInfo[bboxId].regression[1];
		faceRect.y2 = faceInfo[bboxId].bbox.y2 + regw * faceInfo[bboxId].regression[2];
		faceRect.x2 = faceInfo[bboxId].bbox.x2 + regh * faceInfo[bboxId].regression[3];
		faceRect.score = faceInfo[bboxId].bbox.score;

		tempFaceInfo.bbox = faceRect;
		tempFaceInfo.regression = faceInfo[bboxId].regression;
		if (stage == 3)
			tempFaceInfo.facePts = faceInfo[bboxId].facePts;
		bboxes.push_back(tempFaceInfo);
	}
	return bboxes;
}

// compute the padding coordinates (pad the bounding boxes to square)
void MTCNN::Padding(int img_w, int img_h) {
	for (int i = 0; i < regressed_rects_.size(); i++) {
		FaceInfo tempFaceInfo;
		tempFaceInfo = regressed_rects_[i];
		tempFaceInfo.bbox.y2 = (regressed_rects_[i].bbox.y2 >= img_w) ? img_w : regressed_rects_[i].bbox.y2;
		tempFaceInfo.bbox.x2 = (regressed_rects_[i].bbox.x2 >= img_h) ? img_h : regressed_rects_[i].bbox.x2;
		tempFaceInfo.bbox.y1 = (regressed_rects_[i].bbox.y1 < 1) ? 1 : regressed_rects_[i].bbox.y1;
		tempFaceInfo.bbox.x1 = (regressed_rects_[i].bbox.x1 < 1) ? 1 : regressed_rects_[i].bbox.x1;
		regressed_pading_.push_back(tempFaceInfo);
	}
}

void MTCNN::GenerateBoundingBox(Blob<float>* confidence, Blob<float>* reg,
	float scale, float thresh, int image_width, int image_height) {
	int stride = 2;
	int cellSize = 12;

	int curr_feature_map_w_ = std::ceil((image_width - cellSize)*1.0 / stride) + 1;
	int curr_feature_map_h_ = std::ceil((image_height - cellSize)*1.0 / stride) + 1;

	//std::cout << "Feature_map_size:"<< curr_feature_map_w_ <<" "<<curr_feature_map_h_<<std::endl;
	int regOffset = curr_feature_map_w_*curr_feature_map_h_;
	// the first count numbers are confidence of face
	int count = confidence->count() / 2;
	const float* confidence_data = confidence->cpu_data();
	confidence_data += count;
	const float* reg_data = reg->cpu_data();

	condidate_rects_.clear();
	for (int i = 0; i < count; i++) {
		if (*(confidence_data + i) >= thresh) {
			int y = i / curr_feature_map_w_;
			int x = i - curr_feature_map_w_ * y;

			float xTop = (int)((x*stride + 1) / scale);
			float yTop = (int)((y*stride + 1) / scale);
			float xBot = (int)((x*stride + cellSize - 1 + 1) / scale);
			float yBot = (int)((y*stride + cellSize - 1 + 1) / scale);
			FaceRect faceRect;
			faceRect.x1 = xTop;
			faceRect.y1 = yTop;
			faceRect.x2 = xBot;
			faceRect.y2 = yBot;
			faceRect.score = *(confidence_data + i);
			FaceInfo faceInfo;
			faceInfo.bbox = faceRect;
			faceInfo.regression = cv::Vec4f(reg_data[i + 0 * regOffset], reg_data[i + 1 * regOffset], reg_data[i + 2 * regOffset], reg_data[i + 3 * regOffset]);
			condidate_rects_.push_back(faceInfo);
		}
	}
}

MTCNN::MTCNN(const std::string &proto_model_dir) {
#ifdef CPU_ONLY
	Caffe::set_mode(Caffe::CPU);
#else
	Caffe::set_mode(Caffe::GPU);
#endif
	/* Load the network. */
	PNet_.reset(new Net<float>((proto_model_dir + "/det1.prototxt"), TEST));
	PNet_->CopyTrainedLayersFrom(proto_model_dir + "/det1.caffemodel");

	CHECK_EQ(PNet_->num_inputs(), 1) << "Network should have exactly one input.";
	CHECK_EQ(PNet_->num_outputs(), 2) << "Network should have exactly two output, one"
		" is bbox and another is confidence.";

	//RNet_.reset(new Net<float>((proto_model_dir+"/det2.prototxt"), TEST));
	RNet_.reset(new Net<float>((proto_model_dir + "/det2_input.prototxt"), TEST));
	RNet_->CopyTrainedLayersFrom(proto_model_dir + "/det2.caffemodel");

	//  CHECK_EQ(RNet_->num_inputs(), 0) << "Network should have exactly one input.";
	//  CHECK_EQ(RNet_->num_outputs(),3) << "Network should have exactly two output, one"
	//                                     " is bbox and another is confidence.";

	ONet_.reset(new Net<float>((proto_model_dir + "/det3_input.prototxt"), TEST));
	ONet_->CopyTrainedLayersFrom(proto_model_dir + "/det3.caffemodel");

	//  CHECK_EQ(ONet_->num_inputs(), 1) << "Network should have exactly one input.";
	//  CHECK_EQ(ONet_->num_outputs(),3) << "Network should have exactly three output, one"
	//                                     " is bbox and another is confidence.";

	Blob<float>* input_layer;
	input_layer = PNet_->input_blobs()[0];
	num_channels_ = input_layer->channels();
	CHECK(num_channels_ == 3 || num_channels_ == 1) << "Input layer should have 1 or 3 channels.";
}

void MTCNN::WrapInputLayer(std::vector<cv::Mat>* input_channels,
	Blob<float>* input_layer, const int height, const int width) {
	float* input_data = input_layer->mutable_cpu_data();
	for (int i = 0; i < input_layer->channels(); ++i) {
		cv::Mat channel(height, width, CV_32FC1, input_data);
		input_channels->push_back(channel);
		input_data += width * height;
	}
}
// multi test image pass a forward
void MTCNN::ClassifyFace_MulImage(const std::vector<FaceInfo>& regressed_rects, cv::Mat &sample_single,
	boost::shared_ptr<Net<float> >& net, double thresh, char netName) {
	condidate_rects_.clear();

	int numBox = regressed_rects.size();
	std::vector<Datum> datum_vector;

	boost::shared_ptr<MemoryDataLayer<float> > mem_data_layer;
	mem_data_layer = boost::static_pointer_cast<MemoryDataLayer<float>>(net->layers()[0]);
	int input_width = mem_data_layer->width();
	int input_height = mem_data_layer->height();

	// load crop_img data to datum
	for (int i = 0; i < numBox; i++) {
		int pad_top = std::abs(regressed_pading_[i].bbox.x1 - regressed_rects[i].bbox.x1);
		int pad_left = std::abs(regressed_pading_[i].bbox.y1 - regressed_rects[i].bbox.y1);
		int pad_right = std::abs(regressed_pading_[i].bbox.y2 - regressed_rects[i].bbox.y2);
		int pad_bottom = std::abs(regressed_pading_[i].bbox.x2 - regressed_rects[i].bbox.x2);

		cv::Mat crop_img = sample_single(cv::Range(regressed_pading_[i].bbox.y1 - 1, regressed_pading_[i].bbox.y2),
			cv::Range(regressed_pading_[i].bbox.x1 - 1, regressed_pading_[i].bbox.x2));
		cv::copyMakeBorder(crop_img, crop_img, pad_left, pad_right, pad_top, pad_bottom, cv::BORDER_CONSTANT, cv::Scalar(0));

		cv::resize(crop_img, crop_img, cv::Size(input_width, input_height), 0, 0, cv::INTER_AREA);
		crop_img = (crop_img - 127.5)*0.0078125;
		Datum datum;
		CvMatToDatumSignalChannel(crop_img, &datum);
		datum_vector.push_back(datum);
	}
	regressed_pading_.clear();

	/* extract the features and store */
	mem_data_layer->set_batch_size(numBox);
	mem_data_layer->AddDatumVector(datum_vector);
	/* fire the network */
	float no_use_loss = 0;
	net->Forward(&no_use_loss);
	//  CHECK(reinterpret_cast<float*>(crop_img_set.at(0).data) == net->input_blobs()[0]->cpu_data())
	//          << "Input channels are not wrapping the input layer of the network.";

	// return RNet/ONet result
	std::string outPutLayerName = (netName == 'r' ? "conv5-2" : "conv6-2");
	std::string pointsLayerName = "conv6-3";

	const boost::shared_ptr<Blob<float> > reg = net->blob_by_name(outPutLayerName);
	const boost::shared_ptr<Blob<float> > confidence = net->blob_by_name("prob1");
	// ONet points_offset != NULL
	const boost::shared_ptr<Blob<float> > points_offset = net->blob_by_name(pointsLayerName);

	const float* confidence_data = confidence->cpu_data();
	const float* reg_data = reg->cpu_data();

	for (int i = 0; i<numBox; i++) {
		if (*(confidence_data + i * 2 + 1) > thresh) {
			FaceRect faceRect;
			faceRect.x1 = regressed_rects[i].bbox.x1;
			faceRect.y1 = regressed_rects[i].bbox.y1;
			faceRect.x2 = regressed_rects[i].bbox.x2;
			faceRect.y2 = regressed_rects[i].bbox.y2;
			faceRect.score = *(confidence_data + i * 2 + 1);
			FaceInfo faceInfo;
			faceInfo.bbox = faceRect;
			faceInfo.regression = cv::Vec4f(reg_data[4 * i + 0], reg_data[4 * i + 1], reg_data[4 * i + 2], reg_data[4 * i + 3]);

			// x x x x x y y y y y
			if (netName == 'o') {
				FacePts face_pts;
				const float* points_data = points_offset->cpu_data();
				float w = faceRect.y2 - faceRect.y1 + 1;
				float h = faceRect.x2 - faceRect.x1 + 1;
				for (int j = 0; j < 5; j++) {
					face_pts.y[j] = faceRect.y1 + *(points_data + j + 10 * i) * h - 1;
					face_pts.x[j] = faceRect.x1 + *(points_data + j + 5 + 10 * i) * w - 1;
				}
				faceInfo.facePts = face_pts;
			}
			condidate_rects_.push_back(faceInfo);
		}
	}
}
bool MTCNN::CvMatToDatumSignalChannel(const cv::Mat& cv_mat, Datum* datum) {
	if (cv_mat.empty())
		return false;
	int channels = cv_mat.channels();

	datum->set_channels(cv_mat.channels());
	datum->set_height(cv_mat.rows);
	datum->set_width(cv_mat.cols);
	datum->set_label(0);
	datum->clear_data();
	datum->clear_float_data();
	datum->set_encoded(false);

	int datum_height = datum->height();
	int datum_width = datum->width();
	if (channels == 3) {
		for (int c = 0; c < channels; c++) {
			for (int h = 0; h < datum_height; ++h) {
				for (int w = 0; w < datum_width; ++w) {
					const float* ptr = cv_mat.ptr<float>(h);
					datum->add_float_data(ptr[w*channels + c]);
				}
			}
		}
	}

	return true;
}

void MTCNN::Detect(const cv::Mat& image, std::vector<FaceInfo>& faceInfo, int minSize, double* threshold, double factor) {

	// 2~3ms
	// invert to RGB color space and float type
	cv::Mat sample_single, resized;
	image.convertTo(sample_single, CV_32FC3);
	cv::cvtColor(sample_single, sample_single, cv::COLOR_BGR2RGB);
	sample_single = sample_single.t();

	int height = image.rows;
	int width = image.cols;
	int minWH = min(height, width);
	int factor_count = 0;
	double m = 12. / minSize;
	minWH *= m;
	std::vector<double> scales;
	while (minWH >= 12)
	{
		scales.push_back(m * std::pow(factor, factor_count));
		minWH *= factor;
		++factor_count;
	}

	// 11ms main consum
	Blob<float>* input_layer = PNet_->input_blobs()[0];
	for (int i = 0; i < factor_count; i++)
	{
		double scale = scales[i];
		int ws = std::ceil(height*scale);
		int hs = std::ceil(width*scale);

		// wrap image and normalization using INTER_AREA method
		cv::resize(sample_single, resized, cv::Size(ws, hs), 0, 0, cv::INTER_AREA);
		resized.convertTo(resized, CV_32FC3, 0.0078125, -127.5*0.0078125);

		// input data
		input_layer->Reshape(1, 3, hs, ws);
		PNet_->Reshape();
		std::vector<cv::Mat> input_channels;
		WrapInputLayer(&input_channels, PNet_->input_blobs()[0], hs, ws);
		cv::split(resized, input_channels);

		// check data transform right
		CHECK(reinterpret_cast<float*>(input_channels.at(0).data) == PNet_->input_blobs()[0]->cpu_data())
			<< "Input channels are not wrapping the input layer of the network.";
		PNet_->Forward();

		// return result
		Blob<float>* reg = PNet_->output_blobs()[0];
		//const float* reg_data = reg->cpu_data();
		Blob<float>* confidence = PNet_->output_blobs()[1];
		GenerateBoundingBox(confidence, reg, scale, threshold[0], ws, hs);
		std::vector<FaceInfo> bboxes_nms = NonMaximumSuppression(condidate_rects_, 0.5, 'u');
		total_boxes_.insert(total_boxes_.end(), bboxes_nms.begin(), bboxes_nms.end());
	}

	int numBox = total_boxes_.size();
	if (numBox != 0) {
		total_boxes_ = NonMaximumSuppression(total_boxes_, 0.7, 'u');
		regressed_rects_ = BoxRegress(total_boxes_, 1);
		total_boxes_.clear();

		Bbox2Square(regressed_rects_);
		Padding(width, height);

		/// Second stage
		ClassifyFace_MulImage(regressed_rects_, sample_single, RNet_, threshold[1], 'r');
		condidate_rects_ = NonMaximumSuppression(condidate_rects_, 0.7, 'u');
		regressed_rects_ = BoxRegress(condidate_rects_, 2);

		Bbox2Square(regressed_rects_);
		Padding(width, height);

		/// three stage
		numBox = regressed_rects_.size();
		if (numBox != 0) {
			ClassifyFace_MulImage(regressed_rects_, sample_single, ONet_, threshold[2], 'o');
			regressed_rects_ = BoxRegress(condidate_rects_, 3);
			faceInfo = NonMaximumSuppression(regressed_rects_, 0.7, 'm');
		}
	}
	regressed_pading_.clear();
	regressed_rects_.clear();
	condidate_rects_.clear();
}

int main(int argc, char **argv)
{
	::google::InitGoogleLogging(argv[0]);
	double threshold[3] = { 0.6, 0.7, 0.5 };
	double factor = 0.709;
	int minSize = 50;
	string proto_model_dir = "model";

	MTCNN detector(proto_model_dir);

#if FROM_VIDEO
	cv::VideoCapture cap(0);
#else
	string imgname = "test.jpg";
#endif
	cv::Mat frame;
#if FROM_VIDEO
	while (cap.read(frame)) {
#else
		string imageName = imgname;
		frame = cv::imread(imageName);
#endif
		clock_t t1 = clock();
		std::vector<FaceInfo> faceInfo;
		detector.Detect(frame, faceInfo, minSize, threshold, factor);
		std::cout << "Detect " << frame.rows << "X" << frame.cols << " Time Using GPU-CUDNN: " << (clock() - t1)*1.0 / 1000 << std::endl;
		for (int i = 0; i < faceInfo.size(); i++) {
			float x = faceInfo[i].bbox.x1;
			float y = faceInfo[i].bbox.y1;
			float h = faceInfo[i].bbox.x2 - faceInfo[i].bbox.x1 + 1;
			float w = faceInfo[i].bbox.y2 - faceInfo[i].bbox.y1 + 1;
			cv::rectangle(frame, cv::Rect(y, x, w, h), cv::Scalar(255, 0, 0), 2);
		}
		for (int i = 0; i < faceInfo.size(); i++) {
			FacePts facePts = faceInfo[i].facePts;
			for (int j = 0; j < 5; j++)
				cv::circle(frame, cv::Point(facePts.y[j], facePts.x[j]), 3, cv::Scalar(0, 0, 255), 3);
		}
		cv::imshow("img", frame);
#if FROM_VIDEO
		if ((char)cv::waitKey(1) == 'q')
			break;
	}
#else
		string resultpath = resultdir + "/" + imgname;
		cv::imwrite(resultpath, frame);
		cv::waitKey();
#endif
	return 0;
}

相关 [joint face detection] 推荐:

Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Neural Networks | 邹进屹的博客

- -
第三个网络叫ONet,对第二个CNN获得的人脸区域进行再次训练获得是否是人脸,人脸坐标以及五个特征点. 以下项目时MTCNN的具体代码实现. 项目地址:https://github.com/pangyupo/mxnet_mtcnn_face_detection. // MTCNN_VS2015.cpp : 定义控制台应用程序的入口点.

使用JavaScript实现Motion Detection

- - Web App Trend
本篇文章作者 Romuald Quantin,他成为一个前端开发者已经13年有余,在过去的13年中,他一直尝试动画、视频等方面的开发. Romuald在音乐、艺术欣赏、声音设计和音乐方面都颇有造诣,这些使他在开发过程当中受益匪浅,下面我们就来看一下他是如何使用JavaScript实现Motion Detection的吧.

Twitterにグループチャット機能を加えたJoint

- 三十不归 - TechCrunch Japan
約1か月前に、Tom Anderson(またの名Myspace Tom)が、彼が最近好きになったソーシャルネットワークGoogle+について書いた. その中で彼は、Twitterにいくつかアドバイスをしている. 今のTwitterに満足している人は多いと思うが、Andersonは、ユーザ体験のクォリティを上げるためには、あといくつか、簡単な機能を増やせ、と言っている.

Android Ice Cream Sandwich 新增识别脸部解锁功能 -- Face Unlock

- kxxoling - Engadget 中国版
Ice Cream Sandwich 的其中一个新功能就是 -- Face Unlock. 它利用脸部辨识技术进行有关解锁的工作,只是 Google 的 Matias Durate 在现场进行示范时出了一点问题,手机无法辨识主人,未能进行解锁,但这可能只是一时的问题,我们待会动手玩的时候,会多作尝试.

Android Ice Cream Sandwich 新增辨識臉部解鎖功能 -- Face Unlock

- 糜荼 - Engadget 中文版
Ice Cream Sandwich 的其中一個新功能就是 -- Face Unlock. 它利用臉部辨識技術進行有關解鎖的工作,只是 Google 的 Matias Durate 在現場進行示範時出了一點問題,手機無法辨識主人,未能進行解鎖,但這可能只是一時的問題,我們待會動手玩的時候,會多作嘗試.

Andy Rubin:Ice Cream Sandwich 的 Face Unlock 功能由 PittPatt 設計

- SotongDJ - Engadget 中文版
當你仍在為智慧型手機支援臉部辨識解鎖功能而興奮的時候,Google Mobile SVP Andy Rubin 在香港 Asia D: All Things Digital 接受訪問時指出,Face Unlock 正是來自今年年初收購的 PittPatt. 雖然今天早上發表 Android 4.0 Ice Cream Sandwich 時,Face Unlock 一度失敗、無法進行臉部解鎖功能,但 Andy Rubin 在現場再次示範時,成功進行有關解鎖功能,而他也請主持 Walt Mossberg 嘗試用其臉部來解鎖(當然是失敗收場),以確定其準確度,看來我們真的不必擔心太多.

Andy Rubin:Ice Cream Sandwich 的 Face Unlock 功能由 PittPatt 设计

- 云飞风起 - Engadget 中国版
当你仍在为智能型手机支持脸部辨识解锁功能而兴奋的时候,Google Mobile SVP Andy Rubin 在香港 Asia D: All Things Digital 接受访问时指出,Face Unlock 正是来自今年年初收购的 PittPatt. 虽然今天早上发表 Android 4.0 Ice Cream Sandwich 时,Face Unlock 一度失败、无法进行脸部解锁功能,但 Andy Rubin 在现场再次示范时,成功进行有关解锁功能,而他也请主持 Walt Mossberg 尝试用其脸部来解锁(当然是失败收场),以确定其准确度,看来我们真的不必担心太多.

时间序列分段算法 [Time series Breakout Detection]

- - ITeye博客
在时间序列分析中,断点检测(breakout detection)是一个很基本的问题. 通过捕捉时序数据中的断点(breakout),来发现时序数据所表示的系统在过去是否发生了某种事件(event),进而为系统诊断提供必要的数据支持. 为了实现对时序断点的检测,我们首先需要对时序的整体时序做拟合. 这里我们通过一条直线来拟合一段时序,如果时序的趋势发生了变化,则用多条直线来拟合整条时序数据.

Community Detection 社群发现算法 - CSDN博客

- -
        社区发现(Community Detection)算法用来发现网络中的社区结构,也可以视为一种广义的. 以下是我的一个 PPT 报告,分享给大家.         从上述定义可以看出:社区是一个比较含糊的概念,只给出了一个定性的刻画. 另外需要注意的是,社区是一个子图,包含顶点和边.

MobileNetSSD通过Ncnn前向推理框架在Android端的使用--Cmake编译(目标检测 objection detection)补充篇章(多目标也可以显示) - Che_Hongshu - CSDN博客

- -
安装win10+linux-Ubuntu16.04的双系统(超细致). ubuntu16.04+caffe+CUDA10.0+cudnn7.4+opencv2.4.9.1+python2.7 (超超细致). Caffe_ssd安装以及利用VOC2012,VOC2007数据集测试VGG_SSD网络. Caffe实现MobileNetSSD以及各个文件的具体解释,利用自己的数据集dataset训练MobileNetSSD建立模型.