-
Notifications
You must be signed in to change notification settings - Fork 121
/
Copy pathcreate_dataset.py
196 lines (157 loc) · 7.75 KB
/
create_dataset.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
import os
import cv2
import imageio
import numpy as np
import tensorflow as tf
import torch
from pytorch3d import renderer, transforms
from pytorch3d.structures import Meshes
from tqdm import tqdm
import utils
from lib import meshio
from lib.face_segment import Segment
from lib.image_cropper import ImageCropper
from lib.rbf import Shape_Transfer
from lib.uv_creator import UVCreator
class Deep3DFace():
def __init__(self, sess, graph, img_size=224, batch_size=1):
self.sess = sess
self.graph = graph
self.img_size = img_size
self.batch_size = batch_size
self.bfm = utils.BFM_model('.', 'data/models/bfm2009_face.mat')
refer_mesh = meshio.Mesh('data/mesh/bfm09_face.obj')
self.num_bfm_vert = refer_mesh.vertices.shape[0]
bfm_offset = meshio.Mesh('data/mesh/bfm09_face_offset.obj').vertices.astype(
np.float32)
self.bfm_offset = bfm_offset - refer_mesh.vertices.astype(np.float32)
self.vert_mean = np.reshape(self.bfm.shapeMU, [-1, 3])
self.vert_mean += self.bfm_offset
with tf.name_scope('inputs'):
self.ph_images = tf.compat.v1.placeholder(
tf.float32, (self.batch_size, self.img_size, self.img_size, 3),
'input_rgbas')
self.input_images = (self.ph_images + 1) * 127.5
self.infer_bfm()
def infer_bfm(self):
with tf.io.gfile.GFile('data/models/FaceReconModel.pb', 'rb') as f:
face_rec_graph_def = tf.compat.v1.GraphDef()
face_rec_graph_def.ParseFromString(f.read())
def get_emb_coeff(net_name, inputs):
resized = inputs
if self.img_size != 224:
resized = tf.image.resize(inputs, [224, 224])
bgr_inputs = resized[..., ::-1]
tf.import_graph_def(face_rec_graph_def, name=net_name,
input_map={'input_imgs:0': bgr_inputs})
coeff = self.graph.get_tensor_by_name(net_name + '/coeff:0')
return coeff
self.coeff_test = get_emb_coeff('facerec_test', self.input_images)
shape_coef, exp_coef, _, _, _, _ = utils.split_bfm09_coeff(self.coeff_test)
shapePC = tf.constant(self.bfm.shapePC, dtype=tf.float32)
expPC = tf.constant(self.bfm.expressionPC, dtype=tf.float32)
neu_vert = tf.einsum('ij,aj->ai', shapePC, shape_coef)
vertice = neu_vert + tf.einsum('ij,aj->ai', expPC, exp_coef)
neu_vert = tf.reshape(
neu_vert, [self.batch_size, self.num_bfm_vert, 3]) + self.vert_mean
vertice = tf.reshape(
vertice, [self.batch_size, self.num_bfm_vert, 3]) + self.vert_mean
self.vert_test = vertice - tf.reduce_mean(self.vert_mean, axis=0,
keepdims=True)
self.neu_vert_test = neu_vert - tf.reduce_mean(self.vert_mean, axis=0,
keepdims=True)
def predict(self, images, neutral=False):
feed_dict = {self.ph_images: images}
if neutral:
fetches = [self.coeff_test, self.vert_test, self.neu_vert_test]
coeffs, vertices, neu_vert = self.sess.run(fetches, feed_dict)
return coeffs.squeeze(0), vertices.squeeze(0), neu_vert.squeeze(0)
else:
fetches = [self.coeff_test, self.vert_test]
coeffs, vertices = self.sess.run(fetches, feed_dict)
return coeffs.squeeze(0), vertices.squeeze(0)
def main():
device = 'cpu'
device = 'cuda'
img_size = 1024
input_dir = 'data\\dataset\\CelebAMask-HQ\\CelebA-HQ-img'
output_dir = 'data\\dataset\\celeba_hq'
os.makedirs(output_dir, exist_ok=True)
nsh_face_mesh = meshio.Mesh('data/mesh/230/nsh_bfm_face.obj')
cropper = ImageCropper(img_size, use_dlib=False)
segmenter = Segment(device)
transfer = Shape_Transfer('230', device=device)
uv_creator = UVCreator('230', im_size=1024, uv_size=2048, device=device)
R, T = renderer.look_at_view_transform(10, 0, 0)
cameras = renderer.OpenGLPerspectiveCameras(znear=0.001, zfar=30.0,
aspect_ratio=1.0, fov=12.5936,
degrees=True, R=R, T=T,
device=device)
raster_settings = renderer.RasterizationSettings(image_size=1024,
blur_radius=0.0,
faces_per_pixel=1,
bin_size=0)
rasterizer = renderer.MeshRasterizer(cameras=cameras,
raster_settings=raster_settings)
gpu_config = tf.compat.v1.ConfigProto(allow_soft_placement=True,
log_device_placement=False)
# pylint: disable=no-member
gpu_config.gpu_options.allow_growth = True
with tf.compat.v1.Graph().as_default() as graph, tf.compat.v1.device(
'/cpu'), tf.compat.v1.Session(config=gpu_config) as sess:
reconstructor = Deep3DFace(sess, graph, img_size)
# for img_path in tqdm(img_paths):
for index in tqdm(range(30000)):
img_path = os.path.join(input_dir, '{}.jpg'.format(index))
image = cv2.imread(img_path, cv2.IMREAD_UNCHANGED)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
image = utils.center_crop_resize(image, img_size)
image = cropper.crop_image(image)
image = image / 127.5 - 1.0
images = np.expand_dims(image, axis=0).astype(np.float32)
coeff, bfm_vert = reconstructor.predict(images)
_, _, _, [angles], _, [translation] = utils.split_bfm09_coeff(coeff[None])
nsh_vert = transfer.transfer_shape(bfm_vert)
nsh_face_vert = nsh_vert[uv_creator.nsh_face_start_idx:]
transformer = transforms.Transform3d(device=device)
transformer = transformer.rotate_axis_angle(angles[0], 'X', degrees=False)
transformer = transformer.rotate_axis_angle(angles[1], 'Y', degrees=False)
transformer = transformer.rotate_axis_angle(angles[2], 'Z', degrees=False)
transformer = transformer.translate(translation[0], translation[1],
translation[2])
nsh_trans_vert = transformer.transform_points(
torch.from_numpy(nsh_face_vert[None].astype(np.float32)).to(device))
nsh_shift_vert = cameras.get_world_to_view_transform().transform_points(
nsh_trans_vert).data.cpu().numpy()[0] * [-1, 1, -1]
nsh_trans_mesh = Meshes(
nsh_trans_vert,
torch.from_numpy(nsh_face_mesh.triangles[None].astype(
np.int32)).to(device))
fragment = rasterizer(nsh_trans_mesh)
pix_to_face = fragment.pix_to_face.data.cpu().numpy()
visible_face = np.unique(pix_to_face)[1:]
visible_vert = nsh_face_mesh.triangles[visible_face]
visible_vert = np.unique(visible_vert)
vert_alpha = np.zeros([nsh_shift_vert.shape[0], 1])
vert_alpha[visible_vert] = 1
nsh_shift_vert_alpha = np.concatenate([nsh_shift_vert, vert_alpha],
axis=-1)
_, segments = segmenter.segment(images, batch_size=1, all_seg=True)
segment = segments[0]
uv_map = uv_creator.create_nsh_uv_np(nsh_shift_vert_alpha,
utils.to_uint8(image), segment, True)
uv_map = cv2.resize(uv_map, (1024, 1024), interpolation=cv2.INTER_AREA)
uv_map[..., 3] = uv_map[..., 3] + uv_map[..., 4] * 128
np.save(os.path.join(output_dir, '{:>05d}_params.npy'.format(index)),
coeff.astype(np.float32))
image_with_seg = np.concatenate(
[utils.to_uint8(image), 255 - segment[..., None]], axis=-1)
imageio.imwrite(
os.path.join(output_dir, '{:>05d}_image.png'.format(index)),
image_with_seg)
imageio.imwrite(os.path.join(output_dir, '{:>05d}_uv.png'.format(index)),
uv_map[..., :4])
np.save(os.path.join(output_dir, '{:>05d}_nsh_vert.npy'.format(index)),
nsh_face_vert.astype(np.float32))
if __name__ == '__main__':
main()