-
Notifications
You must be signed in to change notification settings - Fork 11
/
demo_in_the_wild.py
197 lines (177 loc) · 6.4 KB
/
demo_in_the_wild.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
import cv2
import numpy as np
import pygame
import torch
import time
import torch.backends.cudnn as cudnn
import pyrealsense2 as rs
import jax.numpy as npj
import PIL.Image as Image
import glob
from jax import grad, jit, vmap
from jax.experimental import optimizers
from torchvision.transforms import functional
import pickle
from manolayer import ManoLayer
from model import HandNetInTheWild
from checkpoints import CheckpointIO
import utils
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
cudnn.benchmark = True
mano_layer = ManoLayer(center_idx=9, side="right", mano_root=".", use_pca=False, flat_hand_mean=True,)
mano_layer = jit(mano_layer)
@jit
def hm_to_kp2d(hm):
b, c, w, h = hm.shape
hm = hm.reshape(b,c,-1)
hm = hm/npj.sum(hm,-1,keepdims=True)
coord_map_x = npj.tile(npj.arange(0,w).reshape(-1,1), (1,h))
coord_map_y = npj.tile(npj.arange(0,h).reshape(1,-1), (w,1))
coord_map_x = coord_map_x.reshape(1,1,-1)
coord_map_y = coord_map_y.reshape(1,1,-1)
x = npj.sum(coord_map_x * hm,-1,keepdims=True)
y = npj.sum(coord_map_y * hm,-1,keepdims=True)
kp_2d = npj.concatenate((y,x),axis=-1)
return kp_2d
@jit
def reinit_scale(joint,kp2d,t):
xy = joint[:,:2]
kp2d = kp2d[0]
s1 = npj.sum(xy*(kp2d-t))
s2 = npj.sum((xy**2))
s = s1/s2
#bone = bone * npj.max([s,0.9])
return s
@jit
def geo(joint):
idx_a = npj.array([1,5,9,13,17])
idx_b = npj.array([2,6,10,14,18])
idx_c = npj.array([3,7,11,15,19])
idx_d = npj.array([4,8,12,16,20])
p_a = joint[:,idx_a,:]
p_b = joint[:,idx_b,:]
p_c = joint[:,idx_c,:]
p_d = joint[:,idx_d,:]
v_ab = p_a - p_b #(B, 5, 3)
v_bc = p_b - p_c #(B, 5, 3)
v_cd = p_c - p_d #(B, 5, 3)
loss_1 = npj.abs(npj.sum(npj.cross(v_ab, v_bc, -1) * v_cd, -1)).mean()
loss_2 = - npj.clip(npj.sum(npj.cross(v_ab, v_bc, -1) * npj.cross(v_bc, v_cd, -1)), -npj.inf, 0).mean()
loss = 10000*loss_1 + 100000*loss_2
return loss
@jit
def residuals(input_list,so3_init,beta_init,kp2d,s,t):
so3 = input_list['so3']
beta = input_list['beta']
so3 = so3[npj.newaxis,...]
beta = beta[npj.newaxis,...]
_, joint_mano, _ = mano_layer(
pose_coeffs = so3,
betas = beta
)
bone_pred = npj.linalg.norm(joint_mano[:, 0, :] - joint_mano[:, 9, :], axis=1, keepdims=True)
bone_pred = bone_pred[:,npj.newaxis,...]
reg = ((so3 - so3_init)**2)
reg_beta = ((beta - beta_init)**2)
joint_mano = joint_mano / bone_pred
geo_reg = geo(joint_mano)
xy = joint_mano[...,:2]
uv = ( xy * s ) + t
errkp = ((uv - kp2d)**2)
err = 0.01*reg.mean() + 0.01*reg_beta.mean() + 1*errkp.mean() + 100*geo_reg.mean()
return err
@jit
def mano_de(so3, beta):
verts_mano, joint_mano, _ = mano_layer(
pose_coeffs = so3[npj.newaxis,...],
betas = beta[npj.newaxis,...]
)
bone_pred = npj.linalg.norm(joint_mano[:, 0, :] - joint_mano[:, 9, :],axis=1, keepdims=True)
bone_pred = bone_pred[:,npj.newaxis,...]
verts_mano = verts_mano / bone_pred
v = verts_mano[0]
return v
@jit
def mano_de_j(so3, beta):
_, joint_mano, _ = mano_layer(
pose_coeffs = so3[npj.newaxis,...],
betas = beta[npj.newaxis,...]
)
bone_pred = npj.linalg.norm(joint_mano[:, 0, :] - joint_mano[:, 9, :],axis=1, keepdims=True)
bone_pred = bone_pred[:,npj.newaxis,...]
joint_mano = joint_mano / bone_pred
j = joint_mano[0]
return j
def live_application():
model = HandNetInTheWild()
model = model.to(device)
checkpoint_io = CheckpointIO('.', model=model)
load_dict = checkpoint_io.load('checkpoints/model.pt')
model.eval()
dd = pickle.load(open("MANO_RIGHT.pkl", 'rb'), encoding='latin1')
face = np.array(dd['f'])
renderer = utils.OpendrRenderer(
img_size=256)
gr = jit(grad(residuals))
lr = 0.03
opt_init, opt_update, get_params = optimizers.adam(lr, b1=0.5, b2=0.5)
opt_init = jit(opt_init)
opt_update = jit(opt_update)
get_params = jit(get_params)
i = 0
img_list = glob.glob("./demo/*")
with torch.no_grad():
for img_path in img_list:
i = i + 1
img = np.array(Image.open(img_path))
if img is None:
continue
if img.shape[0] > img.shape[1]:
margin = int((img.shape[0] - img.shape[1]) / 2)
img = img[margin:-margin]
elif img.shape[0] < img.shape[1]:
margin = int((img.shape[1] - img.shape[0]) / 2)
img = img[:, margin:-margin]
img = cv2.resize(img, (256, 256),cv2.INTER_LINEAR)
frame = img.copy()
img = functional.to_tensor(img).float()
img = functional.normalize(img, [0.5, 0.5, 0.5], [1, 1, 1])
img = img.unsqueeze(0).to(device)
hm, so3, beta, joint_root, bone = model(img)
kp2d = hm_to_kp2d(hm.detach().cpu().numpy())*4
t = kp2d[0,9,:]
so3 = so3[0].detach().cpu().float().numpy()
beta = beta[0].detach().cpu().float().numpy()
bone = bone[0].detach().cpu().numpy()
joint_root = joint_root[0].detach().cpu().numpy()
so3 = npj.array(so3)
beta = npj.array(beta)
bone = npj.array(bone)
joint_root = npj.array(joint_root)
kp2d = npj.array(kp2d)
so3_init = so3
beta_init = beta
joint = mano_de_j(so3, beta)
s = reinit_scale(joint,kp2d,t)
params = {'so3':so3, 'beta':beta}
so3 = params['so3']
beta = params['beta']
opt_state = opt_init(params)
n = 0
while n < 20:
n = n + 1
params = get_params(opt_state)
grads = gr(params,so3_init,beta_init,kp2d,s,t)
opt_state = opt_update(n, grads, opt_state)
joint = mano_de_j(so3, beta)
s = reinit_scale(joint,kp2d,t)
params = get_params(opt_state)
so3 = params['so3']
beta = params['beta']
v = mano_de(so3, beta)
v = v*s + np.array([t[0],t[1],0])
frame1 = renderer.render(v.copy(),face,frame)
cv2.imwrite("./out/" + str(i) + "_input.png", np.flip(frame,-1))
cv2.imwrite("./out/" + str(i) + "_output.png", np.flip(frame1,-1))
if __name__ == '__main__':
live_application()