-
Notifications
You must be signed in to change notification settings - Fork 1
/
facemesh.py
262 lines (197 loc) · 8.89 KB
/
facemesh.py
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
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
import platform
import cv2
import numpy as np
import tensorflow as tf
import tflite_runtime.interpreter as tflite
from postprocessing import nms_oneclass
# EdgeTPU shared lib name
EDGETPU_SHARED_LIB = {
'Linux': 'libedgetpu.so.1',
'Darwin': 'libedgetpu.1.dylib',
'Windows': 'edgetpu.dll'
}[platform.system()]
class BaseInferencer:
def __init__(self, model_path, edgetpu=True):
experimental_delegates = [tf.lite.experimental.load_delegate(EDGETPU_SHARED_LIB)] if edgetpu else None
self.interpreter = tflite.Interpreter(
model_path=model_path,
experimental_delegates=
experimental_delegates)
self.interpreter.allocate_tensors()
self.input_idx = self.interpreter.get_input_details()[0]['index']
self.input_shape = self.interpreter.get_input_details()[0]['shape'][1:3]
def inference(self, src):
raise NotImplementedError("inference not implemented!")
class FaceDetector(BaseInferencer):
SCORE_THRESH = 0.75
MAX_FACE_NUM = 10
ANCHOR_STRIDES = [8, 16]
ANCHOR_NUM = [2, 6]
def __init__(self, model_path, edgetpu=True):
super(FaceDetector, self).__init__(model_path, edgetpu)
self.outputs_idx = {}
for output in self.interpreter.get_output_details():
self.outputs_idx[output['name']] = output['index']
self.anchors = self.create_anchors(self.input_shape)
def inference(self, image):
# todo: input type check
# convert to float32
image_ = cv2.resize(image, tuple(self.input_shape)).astype(np.float32)
image_ = (image_ - 128.0) / 128.0
image_ = image_[None, ...]
# invoke
self.interpreter.set_tensor(self.input_idx, image_)
self.interpreter.invoke()
scores = self.interpreter.get_tensor(self.outputs_idx['classificators']).squeeze()
scores = 1 / (1 + np.exp(-scores))
bboxes = self.interpreter.get_tensor(self.outputs_idx['regressors']).squeeze()
bboxes_decoded, landmarks, scores = self.decode(scores, bboxes)
bboxes_decoded *= image.shape[0]
landmarks *= image.shape[0]
if len(bboxes_decoded) == 0:
return np.array([]), np.array([]), np.array([])
keep_mask = nms_oneclass(bboxes_decoded, scores) # np.ones(pred_bbox.shape[0]).astype(bool)
bboxes_decoded = bboxes_decoded[keep_mask]
landmarks = landmarks[keep_mask]
scores = scores[keep_mask]
return bboxes_decoded, landmarks, scores
def decode(self, scores, bboxes):
w, h = self.input_shape
cls_mask = scores > self.SCORE_THRESH
if cls_mask.sum() == 0:
return np.array([]), np.array([]), np.array([])
scores = scores[cls_mask]
bboxes = bboxes[cls_mask]
bboxes_anchors = self.anchors[cls_mask]
bboxes_decoded = bboxes_anchors.copy()
bboxes_decoded[:, 0] += bboxes[:, 1] # row
bboxes_decoded[:, 1] += bboxes[:, 0] # columns
bboxes_decoded[:, 0] /= h
bboxes_decoded[:, 1] /= w
pred_w = bboxes[:, 2] / w
pred_h = bboxes[:, 3] / h
topleft_x = bboxes_decoded[:, 1] - pred_w * 0.5
topleft_y = bboxes_decoded[:, 0] - pred_h * 0.5
btmright_x = bboxes_decoded[:, 1] + pred_w * 0.5
btmright_y = bboxes_decoded[:, 0] + pred_h * 0.5
pred_bbox = np.stack([topleft_x, topleft_y, btmright_x, btmright_y], axis=-1)
# decode landmarks
landmarks = bboxes[:, 4:]
landmarks[:, 1::2] += bboxes_anchors[:, 0:1]
landmarks[:, ::2] += bboxes_anchors[:, 1:2]
landmarks[:, 1::2] /= h
landmarks[:, ::2] /= w
return pred_bbox, landmarks, scores
@classmethod
def create_anchors(cls, input_shape):
w, h = input_shape
anchors = []
for s, a_num in zip(cls.ANCHOR_STRIDES, cls.ANCHOR_NUM):
gridCols = (w + s - 1) // s
gridRows = (h + s - 1) // s
x, y = np.meshgrid(np.arange(gridRows), np.arange(gridCols))
x, y = x[..., None], y[..., None]
anchor_grid = np.concatenate([y, x], axis=-1)
anchor_grid = np.tile(anchor_grid, (1, 1, a_num))
anchor_grid = s * (anchor_grid.reshape(-1, 2) + 0.5)
anchors.append(anchor_grid)
return np.concatenate(anchors, axis=0)
class FaceMesher(BaseInferencer):
FACE_KEY_NUM = 468
def __init__(self, model_path, edgetpu=True):
super(FaceMesher, self).__init__(model_path, edgetpu)
outputs_idx_tmp = {}
for output in self.interpreter.get_output_details():
outputs_idx_tmp[output['name']] = output['index']
self.outputs_idx = {'landmark': outputs_idx_tmp['conv2d_20'],
'score': outputs_idx_tmp['conv2d_30']}
def inference(self, image):
h, w = self.input_shape
image_ = cv2.resize(image, tuple(self.input_shape)).astype(np.float32)
image_ = (image_ - 128.0) / 128.0
if len(image_.shape) < 4:
image_ = image_[None, ...]
# invoke
self.interpreter.set_tensor(self.input_idx, image_)
self.interpreter.invoke()
landmarks = self.interpreter.get_tensor(self.outputs_idx['landmark'])
scores = self.interpreter.get_tensor(self.outputs_idx['score'])
# postprocessing
landmarks = landmarks.reshape(self.FACE_KEY_NUM, 3)
landmarks[:, 0] /= w
landmarks[:, 1] /= h
landmarks[:, 0] *= image.shape[1]
landmarks[:, 1] *= image.shape[0]
return landmarks, scores
class FaceAligner:
'''reference to https://www.pyimagesearch.com/2017/05/22/face-alignment-with-opencv-and-python/'''
def __init__(self,
desiredLeftEye=(0.35, 0.35),
desiredFaceWidth=192,
desiredFaceHeight=None):
self.desiredLeftEye = desiredLeftEye
self.desiredFaceWidth = desiredFaceWidth
self.desiredFaceHeight = desiredFaceHeight
if self.desiredFaceHeight is None:
self.desiredFaceHeight = self.desiredFaceWidth
def align(self, image, landmarks):
landmarks = landmarks.astype(int).reshape(-1, 2)
# get left and right eye
left_eye = landmarks[1]
right_eye = landmarks[0]
# computer angle
dY = right_eye[1] - left_eye[1]
dX = right_eye[0] - left_eye[0]
angle = np.degrees(np.arctan2(dY, dX)) - 180
# compute the location of right/left eye in new image
desiredRightEyeX = 1 - self.desiredLeftEye[0]
# get the scale based on the distance
dist = np.sqrt(dY ** 2 + dX ** 2)
desired_dist = (desiredRightEyeX - self.desiredLeftEye[0])
desired_dist *= self.desiredFaceWidth
scale = desired_dist / (dist + 1e-6)
# get the center of eyes
eye_center = (left_eye + right_eye) // 2
# get transformation matrix
M = cv2.getRotationMatrix2D(tuple(eye_center), angle, scale)
# align the center
tX = self.desiredFaceWidth * 0.5
tY = self.desiredFaceHeight * self.desiredLeftEye[1]
M[0, 2] += (tX - eye_center[0])
M[1, 2] += (tY - eye_center[1]) # update translation vector
# apply affine transformation
dst_size = (self.desiredFaceWidth, self.desiredFaceHeight)
output = cv2.warpAffine(image, M, dst_size, flags=cv2.INTER_CUBIC)
return output, M
@staticmethod
def inverse(mesh_landmark, M):
M_inverse = cv2.invertAffineTransform(M)
px = (M_inverse[0, 0] * mesh_landmark[:, 0:1] + M_inverse[0, 1] * mesh_landmark[:, 1:2] + M_inverse[0, 2])
py = (M_inverse[1, 0] * mesh_landmark[:, 0:1] + M_inverse[1, 1] * mesh_landmark[:, 1:2] + M_inverse[1, 2])
mesh_landmark_inverse = np.concatenate([px, py, mesh_landmark[:, 2:]], axis=-1)
return mesh_landmark_inverse
class FacePoseDecoder:
FACE_MODEL_3D = np.array([
(-165.0, 170.0, -135.0), # left eye
(165.0, 170.0, -135.0), # right eye
(0.0, 0.0, 0.0), # Nose
(0.0, -150, -110), # mouth
(-330.0, 100.0, -305.0), # left face
(330.0, 100.0, -305.0), # right face
]) / 4.5 # 4.5 is scale factor
def __init__(self, img_size):
self.size = (640, 640)
# Camera internals
self.focal_length = self.size[1]
self.camera_center = (self.size[1] / 2, self.size[0] / 2)
self.camera_matrix = np.array(
[[self.focal_length, 0, self.camera_center[0]],
[0, self.focal_length, self.camera_center[1]],
[0, 0, 1]], dtype="double")
# Assuming no lens distortion
self.dist_coeffs = np.zeros((4, 1))
def solve(self, landmarks):
landmarks = landmarks.astype(np.float).reshape(-1, 2)
(_, rotation_vector, translation_vector) = cv2.solvePnP(
self.FACE_MODEL_3D, landmarks, self.camera_matrix, self.dist_coeffs)
return rotation_vector, translation_vector