3DFaceRecognitionOnRaspberryPI/frontalize.py at master · appstud/3DFaceRecognitionOnRaspberryPI · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
__author__ = 'Douglas'

import scipy.io as scio
import cv2
import numpy as np
np.set_printoptions(formatter={'float_kind': lambda x: "%.4f" % x})
import pdb
import imutils
class ThreeD_Model:
    def __init__(self, path, name):
        self.load_model(path, name)

    def load_model(self, path, name):
        model = scio.loadmat(path)[name]
        self.out_A=np.asmatrix(np.array([[291.0,0,158.0],[0,291.0,117.0],[0,0,1]]))
        #self.out_A=np.asmatrix(np.array([[1.74475624e+03, 0,5.17641501e+02 ],[0,1.74362364e+03,  9.66243685e+02],[0, 0, 1.0]]),dtype='float32')
        #self.distCoeff=np.array([1.49069659e-01,-1.14636172, -2.42752680e-03, -1.30123452e-03,2.64749918e+00])
        self.distCoeff=np.array([0.03885,-0.2265,0.005354,-0.0016,0.1353])
        self.size_U = model['sizeU'][0, 0][0] #1x2
        self.model_TD = np.asarray(model['threedee'][0,0], dtype='float32') #68x3
        self.indbad = model['indbad'][0, 0]#0x1
        self.ref_U = np.asarray(model['refU'][0,0])

    def drawCoordinateSystems(self,extrinsicMatrix,image,_3Dpoints=np.array([[0,0,0],[0.08, 0, 0], [0, -0.08, 0], [0, 0, 0.08]])):

        _2DprojectedPoints=np.matrix(self.out_A)*np.matrix(extrinsicMatrix)*np.transpose(np.matrix(np.hstack((_3Dpoints,np.ones([len(_3Dpoints),1])))))
        _2DprojectedPoints=_2DprojectedPoints/_2DprojectedPoints[2,:]
        cv2.line(image, tuple(_2DprojectedPoints[0:2,0]), tuple(_2DprojectedPoints[0:2,1]), (0,255,0), 3, 0)
        cv2.line(image, tuple(_2DprojectedPoints[0:2,0]), tuple(_2DprojectedPoints[0:2,2]),(255,0,0),3, 0)
        cv2.line(image, tuple(_2DprojectedPoints[0:2,0]), tuple(_2DprojectedPoints[0:2,3]), (0,0,255),3, 0)
        cv2.putText(image,'X' ,tuple(_2DprojectedPoints[0:2,1]),cv2.FONT_HERSHEY_SIMPLEX,0.4,(0,0,0))
        cv2.putText(image,'Y',tuple(_2DprojectedPoints[0:2,2]),cv2.FONT_HERSHEY_SIMPLEX,0.4,(0,0,0))
        cv2.putText(image,'Z', tuple(_2DprojectedPoints[0:2,3]),cv2.FONT_HERSHEY_SIMPLEX,0.4,(0,0,0))
        return image,_2DprojectedPoints

def frontalize(img, proj_matrix, ref_U, eyemask):
    ACC_CONST = 100
    img = img.astype('float32')
    #print(("query image shape:", img.shape))
    #width,height=140,140


    ###bgind = np.sum(np.abs(ref_U), 2) == 0
    ####try this####
    ###################################################bgind = np.abs(ref_U[:,:,1]) < 60
    ## bgind = np.abs(ref_U[:,:,1]) < 50
    ###bgind = np.abs(ref_U[:,:,1]) < 7
    #bgind=np.zeros([ref_U.shape[0],ref_U.shape[1]])
    #bgind[int(160-width/2):int(160+width/2),int(160-height/2):int(160+height/2)]=1
    #bgind=bgind==0
    # count the number of times each pixel in the query is accessed
    #ref_U =imutils.resize(ref_U[52:250,91:225,:],height=48)
    #eyemask =imutils.resize(eyemask[52:250,91:225],height=48)
    bgind = np.abs(ref_U[:,:,1]) < 60
    threedee = np.reshape(ref_U, (-1, 3), order='F').transpose()

    temp_proj = proj_matrix * np.vstack((threedee, np.ones((1, threedee.shape[1]))))
    temp_proj2 = np.divide(temp_proj[0:2, :], np.tile(temp_proj[2, :], (2,1)))

    bad = np.logical_or(temp_proj2.min(axis=0) < 1, temp_proj2[1, :] > img.shape[0])
    bad = np.logical_or(bad, temp_proj2[0, :] > img.shape[1])
    bad = np.logical_or(bad, bgind.reshape((-1), order='F'))
    bad = np.asarray(bad).reshape((-1), order='F')

    temp_proj2 -= 1

    badind = np.nonzero(bad > 0)[0]
    temp_proj2[:,badind] = 0

    ind = np.ravel_multi_index((np.asarray(temp_proj2[1, :].round(), dtype='int64'), np.asarray(temp_proj2[0, :].round(), dtype='int64')), dims=img.shape[:-1], order='F')

    synth_frontal_acc = np.zeros(ref_U.shape[:-1])
    ind_frontal = np.arange(0, ref_U.shape[0]*ref_U.shape[1])
    rer=60
    c, ic = np.unique(ind, return_inverse=True)
    bin_edges = np.r_[-np.Inf, 0.5 * (c[:-1] + c[1:]), np.Inf]
    count, bin_edges = np.histogram(ind, bin_edges)
    synth_frontal_acc = synth_frontal_acc.reshape(-1, order='F')
    synth_frontal_acc[ind_frontal] = count[ic]
    synth_frontal_acc = synth_frontal_acc.reshape(( rer,-1), order='F')
    ############synth_frontal_acc[bgind] = 0
    #synth_frontal_acc = cv2.GaussianBlur(synth_frontal_acc, (15, 15), 30., borderType=cv2.BORDER_REPLICATE)
    synth_frontal_acc = cv2.GaussianBlur(synth_frontal_acc, (11, 11), 20, borderType=cv2.BORDER_REPLICATE)

    #remap
    mapX = temp_proj2[0,:].astype(np.float32)
    mapY = temp_proj2[1,:].astype(np.float32)

    mapX = np.reshape(mapX,(rer,-1), order = 'F')
    mapY = np.reshape(mapY,(rer,-1), order = 'F')

    frontal_raw = cv2.remap(img, mapX, mapY, cv2.INTER_CUBIC)

    frontal_raw = frontal_raw.reshape((-1,3), order = 'F')
    frontal_raw[badind,:] = 0
    frontal_raw = frontal_raw.reshape((rer,-1, 3), order = 'F')

    # which side has more occlusions?
    midcolumn = int(np.round(ref_U.shape[1]/2))
    sumaccs = synth_frontal_acc.sum(axis=0)
    sum_left = sumaccs[0:midcolumn].sum()
    sum_right = sumaccs[midcolumn+1:].sum()
    sum_diff = sum_left - sum_right

    if np.abs(sum_diff) > ACC_CONST: # one side is ocluded
        ones = np.ones((ref_U.shape[0], midcolumn))
        zeros = np.zeros((ref_U.shape[0], midcolumn))
        if sum_diff > ACC_CONST: # left side of face has more occlusions
            weights = np.hstack((zeros, ones))
        else: # right side of face has more occlusions
            weights = np.hstack((ones, zeros))
        weights = cv2.GaussianBlur(weights, (3, 3), 8, borderType=cv2.BORDER_REPLICATE)

        # apply soft symmetry to use whatever parts are visible in ocluded side
        synth_frontal_acc /= synth_frontal_acc.max()
        weight_take_from_org = 1. / np.exp(0.5+synth_frontal_acc)
        #weight_take_from_org = np.zeros(synth_frontal_acc.shape)
        weight_take_from_sym = 1 - weight_take_from_org

        weight_take_from_org = np.multiply(weight_take_from_org, np.fliplr(weights))
        weight_take_from_sym = np.multiply(weight_take_from_sym, np.fliplr(weights))

        weight_take_from_org = np.tile(weight_take_from_org.reshape(rer,-1, 1), (1, 1, 3))
        weight_take_from_sym = np.tile(weight_take_from_sym.reshape(rer,-1, 1), (1, 1, 3))
        weights = np.tile(weights.reshape(rer,-1, 1), (1, 1, 3))

        denominator = weights + weight_take_from_org + weight_take_from_sym
        frontal_sym = np.multiply(frontal_raw, weights) + np.multiply(frontal_raw, weight_take_from_org) + np.multiply(np.fliplr(frontal_raw), weight_take_from_sym)
        frontal_sym = np.divide(frontal_sym, denominator)

        # exclude eyes from symmetry
        frontal_sym = np.multiply(frontal_sym, 1-eyemask) + np.multiply(frontal_raw, eyemask)
        frontal_raw[frontal_raw > 255] = 255
        frontal_raw[frontal_raw < 0] = 0
        frontal_raw = frontal_raw.astype('uint8')
        frontal_sym[frontal_sym > 255] = 255
        frontal_sym[frontal_sym < 0] = 0
        frontal_sym = frontal_sym.astype('uint8')
    else: # both sides are occluded pretty much to the same extent -- do not use symmetry
        frontal_sym = frontal_raw


    return frontal_raw.astype('uint8'), frontal_sym.astype('uint8')