-
Notifications
You must be signed in to change notification settings - Fork 0
/
step1_recon_3d_face.py
executable file
·141 lines (123 loc) · 6.7 KB
/
step1_recon_3d_face.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
import argparse
import tensorflow as tf
import numpy as np
import cv2
from PIL import Image
import os
from tqdm import tqdm
from scipy.io import loadmat,savemat
from utils.preprocess import POS, headrecon_preprocess_withmask, facerecon_preprocess_yu_5p, facerecon_preprocess
from utils.loader import load_data, load_lm3d, load_center3d, read_facemodel
from utils.recon_face import compute_center2d, compute_faceshape
from utils.create_renderer import create_renderer_graph
from PIL import Image
def load_facerecon_graph(graph_filename):
with tf.gfile.GFile(graph_filename,'rb') as f:
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())
with tf.Graph().as_default() as graph:
input = tf.placeholder(name='input_imgs', shape=[
None, 224, 224, 3], dtype=tf.float32)
tf.import_graph_def(graph_def, name='resnet', input_map={
'input_imgs:0': input})
output = graph.get_tensor_by_name('resnet/coeff:0')
return graph, input, output
def face_recon(input_path, output_path, vis_path=None, s_factor=1.5, focal=1015, center=112, align_nums=10):
print(f'[INFO] [Step1] Read images from {input_path}')
# load BFM
facemodel = read_facemodel()
# read standard landmarks for face recon preprocessing
lm3D = load_lm3d(align_nums)
# read head center for depth recon preprocessing
head_center3d = load_center3d()
# create face recon graph
face_recon_graph, images, coef = load_facerecon_graph('model/model_mask3_white_light.pb')
face_recon_sess = tf.Session(graph=face_recon_graph)
# create renderer graph
depth_render_graph, input_focal, input_center, input_depth, \
input_vertex, input_tri, output_depthmap = create_renderer_graph()
render_sess = tf.Session(graph=depth_render_graph)
imgs_path = [os.path.join(input_path, i) for i in os.listdir(input_path)
if i.endswith('png') or i.endswith('jpg') or i.endswith('jpeg')]
imgs_path = tqdm(imgs_path)
for i, name in enumerate(imgs_path):
# print(i, name.split(os.path.sep)[-1].split('.')[0])
mask_mat_path = os.path.join(input_path, name.split(os.path.sep)[-1].split('.')[0] + '.mat')
if not os.path.exists(mask_mat_path): continue
mask = loadmat(mask_mat_path)['mask']
## load images and corresponding 5 facial landmarks
if align_nums == 5:
img, lm = load_data(name,
os.path.join(input_path, name.split(os.path.sep)[-1].split('.')[0] + '_detection.txt'))
lm = lm[-10:].reshape([5, 2])
input_img, inv_params = facerecon_preprocess_yu_5p(img, lm, lm3D)
elif align_nums == 10:
landmark_path = os.path.join(input_path, name.split(os.path.sep)[-1].split('.')[0] + '_landmark.txt')
if not os.path.exists(landmark_path) : continue
img, lm = load_data(name, landmark_path)
lm = lm.reshape([68, 2])
input_img, inv_params = facerecon_preprocess(img, lm, lm3D)
# recon face
coeff = face_recon_sess.run(coef, feed_dict={images: np.expand_dims(input_img, 0)})[..., :-1]
# preprocess input image for depth recon net
# reproject the reconstructed face to raw image with adjusted focal and center
f = focal * inv_params[0]
p_center = inv_params[0] * center + inv_params[1]
face_shape, face_projection, landmarks_2d = compute_faceshape(coeff, facemodel, inv_params)
# crop the raw image with head center as the image center
center2d, displacement = compute_center2d(head_center3d, coeff, facemodel, f, p_center)
_, s = POS(face_projection.transpose(), facemodel.meanshape.reshape([-1, 3]).transpose())
crop_img, crop_mask, inv_params_, crop_lm, crop_param = headrecon_preprocess_withmask(img, mask, landmarks_2d, center2d.reshape([2]), s*s_factor/100)
# save processed data
data = np.zeros([3 + 257 + 136])
data[0] = f / inv_params_[0]
data[1: 3] = (p_center - inv_params_[1].reshape([2]))/inv_params_[0]
data[3: 260] = coeff.reshape([257])
data[257: 260] = data[257: 260] - displacement.reshape([3])
data[260:] = crop_lm.reshape([136])
face_projection_cropped, _ = compute_center2d(np.expand_dims(face_shape, 0),
np.expand_dims(data[3:260], 0), facemodel, data[0], data[1:3], displace_flag=False, apply_pose=False)
# render face depth
d = 10 - face_shape[:, 2:]
d = np.tile(np.expand_dims(d, 0), [1, 1, 3])
d_map = render_sess.run(output_depthmap, feed_dict={
input_focal: data[0].reshape([1]),
input_center: data[1: 3].reshape([1, 1, 2]),
input_depth: d,
input_vertex: np.expand_dims(face_shape, 0),
input_tri: np.expand_dims(facemodel.tri, 0) - 1 # start from 0
})
if vis_path:
cv2.imwrite(os.path.join(vis_path, name.split(os.path.sep)[-1].split('.')[0]+ '.png'),
crop_img.astype(np.uint8))
cv2.imwrite(os.path.join(vis_path, name.split(os.path.sep)[-1].split('.')[0]+ '_dmap.png'), d_map[0] * 255)
savemat(os.path.join(output_path, name.split(os.path.sep)[-1].split('.')[0]+ '.mat'),
{'img': crop_img.astype(np.uint8),
'mask': crop_mask.astype(np.uint8),
'crop_param': crop_param.astype(np.float32),
'face3d': data.astype(np.float32),
# 0: focal; [1, 3) center; [3,260): face coeff; [260~396): landmark
'face_shape':face_shape.astype(np.float32),
'face_projection': face_projection_cropped.squeeze(0).astype(np.float32),
'face_depthmap': d_map[..., 0].squeeze(0),
'face_mask': d_map[..., -1].squeeze(0),
'face_tri': facemodel.tri}, do_compression=True)
face_recon_sess.close()
render_sess.close()
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--root_dir', default='.')
parser.add_argument('--input_path', default='examples')
parser.add_argument('--save_path', default='output/step1')
parser.add_argument('--vis_path', default=None) # e.g. 'output/step1/vis
# prepare directory
args = parser.parse_args()
input_path = os.path.join(args.root_dir, args.input_path)
save_path = os.path.join(args.root_dir, args.save_path)
vis_path = os.path.join(args.vis_path, args.vis_path) if args.vis_path else None
if not os.path.isdir(save_path):
os.makedirs(save_path)
if vis_path and not os.path.isdir(vis_path):
os.makedirs(vis_path)
# recon 3d face and prepare the input to depth recon
face_recon(input_path, save_path, vis_path)