-
Notifications
You must be signed in to change notification settings - Fork 10
/
Copy patheval(TTA).py
200 lines (196 loc) · 10.4 KB
/
eval(TTA).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
import torch
import torch.optim
import numpy as np
import argparse
import torch.nn as nn
import model
import os
import cv2
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
def dt(a):
return cv2.distanceTransform((a * 255).astype(np.uint8), cv2.DIST_L2, 0)
def trimap_transform(trimap):
h, w = trimap.shape[0], trimap.shape[1]
clicks = np.zeros((h, w, 6))
for k in range(2):
if (np.count_nonzero(trimap[:, :, k]) > 0):
dt_mask = -dt(1 - trimap[:, :, k]) ** 2
L = 320
clicks[:, :, 3 * k] = np.exp(dt_mask / (2 * ((0.02 * L) ** 2)))
clicks[:, :, 3 * k + 1] = np.exp(dt_mask / (2 * ((0.08 * L) ** 2)))
clicks[:, :, 3 * k + 2] = np.exp(dt_mask / (2 * ((0.16 * L) ** 2)))
return clicks
def genwmap(qsss,www):
h,w=qsss+www,qsss+www
www1=np.ones([h,w],np.float32)*www
ws=np.zeros([h,w],np.float32)
for xx in range(h):
for yy in range(w):
ws[xx,yy]=min(xx,yy,h-xx-1,w-yy-1)+1
ws=ws/(www1+1)
return ws
if __name__ == '__main__':
IMG_SCALE = 1. / 255
IMG_MEAN = np.array([0.485, 0.456, 0.406]).reshape((1, 1, 3))
IMG_STD = np.array([0.229, 0.224, 0.225]).reshape((1, 1, 3))
segmodel = model.LFPNet()
segmodel.load_state_dict(torch.load('model.pth',map_location='cpu'),strict=False)
segmodel=segmodel.cuda()
segmodel.eval()
for idx,x in enumerate(os.listdir('./merged/')):
img=cv2.imread('./merged/'+x)
raw_h,raw_w,_=img.shape
alphaall=np.zeros((raw_h,raw_w),np.float32)
img = cv2.imread('./merged/' + x)
h1_, w1_, _ = img.shape
psss=512
qsss=1024-512
wsss=qsss+psss
wsss2=wsss//2
wx=(w1_-1-psss) //qsss+1
hx=(h1_-1-psss) //qsss+1
alphalist=[]
newh=hx*qsss+psss
neww=wx*qsss+psss
raw_h, raw_w, _ = img.shape
imgzys=img.copy()
alls=[]
alltri=[]
ph1=(newh-raw_h)//2
ph2=newh-raw_h-ph1
pw1=(neww-raw_w)//2
pw2=neww-raw_w-pw1
allp=set(list(np.arange(0,16)))
tp = np.zeros((newh, neww,16), np.float32)
wp = np.zeros((newh, neww, 16), np.float32)
apPP = np.zeros((newh, neww, 16), np.float32)
allp = set(list(np.arange(0, 16)))
wsm = genwmap(qsss, psss)
for x11 in [0]:
for y11 in [0]:
pwh=[0,0,0,0]
pwh=[ph1,ph2,pw1,pw2]
imgbp = cv2.copyMakeBorder(imgzys, pwh[0], pwh[1], pwh[2], pwh[3], cv2.BORDER_REPLICATE)
tri = cv2.imread('./trimap/' + x, cv2.IMREAD_GRAYSCALE)
trizys=tri.copy()
tribp = cv2.copyMakeBorder(tri, pwh[0], pwh[1], pwh[2], pwh[3], cv2.BORDER_CONSTANT)
imgbpp=cv2.copyMakeBorder(imgbp, wsss2,wsss2,wsss2,wsss2, cv2.BORDER_REPLICATE)
tribpp=cv2.copyMakeBorder(tribp, wsss2,wsss2,wsss2,wsss2, cv2.BORDER_CONSTANT)
alls = []
for px in range(hx):
for py in range(wx):
nnns=0
for zzz in range(16):
if np.sum(tp[px*qsss:px*qsss+wsss,py*qsss:py*qsss+wsss,zzz])==0:
nnns=zzz
break
wp[px*qsss: px*qsss+wsss, py*qsss:py*qsss+wsss,nnns]=wsm
tp[px*qsss:px*qsss+wsss,py*qsss:py*qsss+wsss,nnns] = 1
alls = []
for fl1, fl2 in zip([0, -1], [0, -1]):
if fl1 >= 0:
imgf = imgbp[px * qsss:px * qsss + wsss, py * qsss:py * qsss + wsss].copy()
trif = tribp[px * qsss:px * qsss + wsss, py * qsss:py * qsss + wsss].copy()
imgrr = cv2.flip(imgf, fl1)
trirr = cv2.flip(trif, fl1)
imgss2=cv2.flip(imgzys.copy(), fl1)
triss2=cv2.flip(trizys.copy(), fl1)
img4x2 = imgbpp[px*qsss:(px)*qsss+wsss+wsss,py*qsss:(py)*qsss+wsss+wsss].copy()
tri4x2 = tribpp[px*qsss:(px)*qsss+wsss+wsss,py*qsss:(py)*qsss+wsss+wsss].copy()
img4x2 = cv2.flip(img4x2, fl1)
tri4x2 = cv2.flip(tri4x2, fl1)
else:
imgf = imgbp[px * qsss:px * qsss + wsss, py * qsss:py * qsss + wsss].copy()
trif = tribp[px * qsss:px * qsss + wsss, py * qsss:py * qsss + wsss].copy()
imgrr = imgf.copy()
trirr = trif.copy()
imgss2 = imgzys.copy()
triss2 = trizys.copy()
img4x2 = imgbpp[px*qsss:(px)*qsss+wsss+wsss,py*qsss:(py)*qsss+wsss+wsss].copy()
tri4x2 = tribpp[px*qsss:(px)*qsss+wsss+wsss,py*qsss:(py)*qsss+wsss+wsss].copy()
for fr1, fr2 in zip([0, 1, 2, -1], [2, 1, 0, -1]):
if fr1 >= 0:
imgrr3 = cv2.rotate(imgrr.copy(), fr1)
trirr3 = cv2.rotate(trirr.copy(), fr1)
imgss3 = cv2.rotate(imgss2.copy(), fr1)
triss3 = cv2.rotate(triss2.copy(), fr1)
img4x3 = cv2.rotate(img4x2.copy(), fr1)
tri4x3 = cv2.rotate(tri4x2.copy(), fr1)
else:
imgrr3= imgrr.copy()
trirr3= trirr.copy()
imgss3= imgss2.copy()
triss3= triss2.copy()
img4x3=img4x2.copy()
tri4x3=tri4x2.copy()
img=imgrr3
tri=trirr3
imgss=imgss3
triss=triss3
img4x=img4x3
tri4x=tri4x3
tritemp = np.zeros([*tri.shape, 2], np.float32)
tritemp[:, :, 0] = (tri == 0)
tritemp[:, :, 1] = (tri == 255)
sixc = trimap_transform(tritemp)
sixc = np.transpose(sixc, [2, 0, 1])
tritemp = np.zeros([*tri4x.shape, 2], np.float32)
tritemp[:, :, 0] = (tri4x == 0)
tritemp[:, :, 1] = (tri4x == 255)
sixc4x = trimap_transform(tritemp)
sixc4x = np.transpose(sixc4x, [2, 0, 1])
h_, w_ = tri.shape
tri2 = np.array(tri, np.float32) / 255.
tri2 = tri2[np.newaxis, np.newaxis, :, :]
tri24x = np.array(tri4x, np.float32) / 255.
tri24x = tri2[np.newaxis, np.newaxis, :, :]
mattinginput = ((img[:, :, ::-1] / 255.) - IMG_MEAN) / IMG_STD
mattinginput4x = ((img4x[:, :, ::-1] / 255.) - IMG_MEAN) / IMG_STD
raw = img[:, :, ::-1] / 255.
raw = np.transpose(raw, [2, 0, 1])[None, :, :, :]
h_, w_ = tri.shape
trixs = np.zeros((1, 3, h_, w_), np.float32)
trixs[0, 0] = (tri == 0)
trixs[0, 1] = (tri == 128)
trixs[0, 2] = (tri == 255)
ntrixs = trixs
trixs = torch.from_numpy(trixs)
h_, w_ = tri4x.shape
tris4x = np.zeros((1, 3, h_, w_), np.float32)
tris4x[0, 0] = (tri4x == 0)
tris4x[0, 1] = (tri4x == 128)
tris4x[0, 2] = (tri4x == 255)
tris4x = torch.from_numpy(tris4x)
mattinginput = np.transpose(mattinginput, (2, 0, 1)).astype(np.float32)
mattinginput = np.array(mattinginput)[np.newaxis, :, :, :]
mattinginput4x = np.transpose(mattinginput4x, (2, 0, 1)).astype(np.float32)
mattinginput4x = np.array(mattinginput4x)[np.newaxis, :, :, :]
i1 = torch.from_numpy(mattinginput)
i2 = torch.from_numpy(mattinginput4x)
sixc = sixc[None, :, :, :]
sixc4x = sixc4x[None, :, :, :]
sixc2 = torch.from_numpy(sixc)
sixc24x = torch.from_numpy(sixc4x)
i2i = torch.cat([i2, tris4x, sixc24x], 1).float().cuda()
i1i = torch.cat([i1, trixs, sixc2], 1).float().cuda()
rawi = torch.from_numpy(raw).float().cuda()
with torch.no_grad():
preda= segmodel(i1i, i2i, rawi)
preda = preda.detach().cpu()
ap = preda[0, 0].numpy() * ntrixs[0, 1] + ntrixs[0, 2]
a1 = ap
if fr2 >= 0:
a1 = cv2.rotate(a1, fr2)
if fl2 >= 0:
a1 = cv2.flip(a1, fl2)
alls.append(a1)
a1 = np.array(sum(alls) * 255. / len(alls))
alls=[]
apPP[px * qsss:px * qsss + wsss, py * qsss:py * qsss + wsss,nnns] = a1
palpha=np.sum(apPP*wp,2)/ np.sum(wp,2)
palpha=np.clip(palpha,0,255)
palpha=np.array(palpha,np.uint8)
wholealpha=palpha[ph1:ph1+raw_h,pw1:pw1+raw_w]
wholealpha[trizys==0]=0
wholealpha[trizys==255]=255
cv2.imwrite('./alpha/' + x, wholealpha)