-
Notifications
You must be signed in to change notification settings - Fork 25
/
demo.py
162 lines (135 loc) · 6.19 KB
/
demo.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
import torch
import numpy as np
import imageio
import config
from easydict import EasyDict
from transformers.texformer import Texformer
from RSC_net.ra_test import RaRunner
from NMR.neural_render_test import NrTextureRenderer
from utils.scipy_deprecated import imresize
import matplotlib.pyplot as plt
class Demo:
def __init__(self):
self.device = 'cuda'
self.opts = EasyDict(src_ch=4, tgt_ch=3, feat_dim=128,
nhead=8, mask_fusion=1, out_ch=3,
checkpoint_path='./pretrained/texformer_ep500.pt')
self.checkpoint_path = self.opts.checkpoint_path
self.model = Texformer(self.opts)
self.model.load_state_dict(torch.load(self.checkpoint_path))
self.model.eval()
self.model.to(self.device)
self.tgt = torch.from_numpy(np.load(config.uv_encoding_path)).permute(2, 0, 1)[None]
self.tgt = (self.tgt * 2 -1).float().to(self.device)
# others
self.rsc_runner = RaRunner()
self.renderer_seg = NrTextureRenderer(224)
self.renderer = NrTextureRenderer(224, factor=0.8)
self.smpl_part_seg_mapping = torch.tensor([0, 3, 3, 1, 5, 5, 2, 4, 4, 6, 6, 7, 7]).long()
def preprocess_img(self, img):
img = img / 255.
img_tensor = torch.from_numpy(img).float().permute(2, 0, 1).unsqueeze(0)
return img_tensor
def get_smpl(self, img_tensor):
scale = self.rsc_runner.get_scale(128)
with torch.no_grad():
# img_tensor is sent into GPU in runner
pred_vertices, pred_cam_t, pred_rotmat, pred_betas = self.rsc_runner.get_3D_batch(img_tensor, scale, pre_process=True)
self.pred_vertices = pred_vertices
self.pred_cam_t = pred_cam_t
def get_segmentation(self):
parts = self.renderer_seg.render_part(self.pred_vertices, self.pred_cam_t, crop_width=None)
parts = parts.cpu().long()
parts = self.smpl_part_seg_mapping[parts]
return parts
def resize_224(self, img):
h, w = img.shape[0:2]
h_target = 224
ratio = h_target / h
w_target = int(ratio * w)
img = imresize(img, [h_target, w_target])
diff = h_target - w_target
before = abs(diff) // 2
after = diff - before
if diff > 0:
# padding
img = np.pad(img, [(0, 0), (before, after), (0, 0)])
elif diff < 0:
# crop
img = img[:, before:- after]
return img
def preprocess_for_texture(self, img_tensor, parts):
parts = parts.unsqueeze(1).float() / 7.0 # [1, 1, 224, 224], float, [0, 1]
h, w = parts.shape[2:]
h_target = 128
w_target = h_target * w / h
if h != 128:
parts = torch.nn.functional.interpolate(parts, [int(h_target), int(w_target)], mode='nearest')
if w_target > 64:
start = int((w_target-64) // 2)
parts = parts[:, :, :, start:start+64]
elif w_target < 64:
before = (64-w_target) // 2
after = 64-w_target-before
parts = torch.nn.functional.pad(parts, [before, after])
parts = parts * 2 - 1
parts = parts.to(self.device)
img_tensor = img_tensor * 2 - 1
img_tensor = torch.nn.functional.interpolate(img_tensor, [128, 128])
img_tensor = img_tensor[:, :, :, 32:32+64].to(self.device)
return img_tensor, parts
def get_coord(self, shape):
y = np.linspace(-1.0, 1.0, num=shape[0])
x = np.linspace(-1.0, 1.0, num=shape[1])
coord_y, coord_x = np.meshgrid(y, x, indexing='ij')
coord = np.concatenate((coord_y[None], coord_x[None]), axis=0)
return torch.from_numpy(coord).float()
def get_texture(self, img, seg):
coord = self.get_coord([128, 64]).unsqueeze(0).to(self.device)
value = torch.cat([coord, img], dim=1)
src = torch.cat([img, seg], dim=1)
out = self.model(self.tgt, src, value)
combine_mask = out[2]
tex_flow = out[0]
uvmap_flow = torch.nn.functional.grid_sample(img, tex_flow.permute(0, 2, 3, 1))
uvmap_rgb = out[1]
uvmap = uvmap_flow * combine_mask + uvmap_rgb * (1-combine_mask)
return uvmap
@torch.no_grad()
def run_demo(self, args):
img_path = args.img_path
img = imageio.imread(img_path)
img_224 = self.resize_224(img)
img_tensor = self.preprocess_img(img_224)
self.get_smpl(img_tensor)
seg_path = args.seg_path
if seg_path is not None:
seg = imageio.imread(seg_path)
parts = torch.from_numpy(seg)[None].long()
else:
parts = self.get_segmentation()
img_tensor, parts = self.preprocess_for_texture(img_tensor, parts) # ~[-1, 1]
uvmap = self.get_texture(img_tensor, parts)
uvmap = (uvmap + 1) / 2
uvmap = uvmap.clamp(0, 1)
rendered_img, _, _ = self.renderer.render(self.pred_vertices, self.pred_cam_t, uvmap)
rendered_img = rendered_img[0].cpu().permute(1, 2, 0).numpy()
rendered_img_rot, _, _ = self.renderer.render(self.pred_vertices, self.pred_cam_t, uvmap, euler=[0, 180, 0])
rendered_img_rot = rendered_img_rot[0].cpu().permute(1, 2, 0).numpy()
figure, axes = plt.subplots(1, 3, figsize=(20, 6))
axes[0].imshow(img); axes[0].axis('off'); axes[0].set_title('input')
axes[1].imshow(rendered_img); axes[1].axis('off'); axes[1].set_title('rendered')
axes[2].imshow(rendered_img_rot); axes[2].axis('off'); axes[2].set_title('rotated')
if args.save:
plt.savefig('demo_imgs/output.png')
else:
plt.show()
if __name__ == '__main__':
from argparse import ArgumentParser
parser = ArgumentParser()
parser.add_argument('--img_path', required=True, help='Please specify the input image path')
parser.add_argument('--seg_path', type=str, default=None, help='Human part segmentation path. If None, use the result of RSC-Net instead')
parser.add_argument('--save', action='store_true', default=False, help='Whether save the output figure?')
args = parser.parse_args()
demo = Demo()
demo.run_demo(args)