-
Notifications
You must be signed in to change notification settings - Fork 2
/
BuildSPInst_A.py
183 lines (172 loc) · 8.92 KB
/
BuildSPInst_A.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
import numpy as np
from LoadData import *
import numpy.matlib
class GetInst_A(object):
def __init__(self, useful_sp_lab, img3d, gt, trpos):
self.useful_sp_lab = [x[0] for x in useful_sp_lab]
self.scale=len(self.useful_sp_lab)
self.ReduceZeros(gt)
self.img3d = img3d
[self.r, self.c, self.l] = np.shape(img3d)
self.num_classes = int(np.max(gt))
self.img2d = np.reshape(img3d,[self.r*self.c, self.l])
self.sp_num = np.array([ np.max(x) for x in self.useful_sp_lab], dtype='int')
gt = np.array(gt, dtype='int')
self.gt1d = np.reshape(gt, [self.r*self.c])
self.gt_tr = np.array(np.zeros([self.r*self.c]), dtype='int')
self.gt_te = self.gt1d
trpos = np.array(trpos, dtype='int')
self.trpos = (trpos[:,0]-1)*self.c+trpos[:,1]-1
###
self.sp_mean = [np.zeros([x, self.l]) for x in self.sp_num]
self.sp_center_px = [np.zeros([x, self.l]) for x in self.sp_num]
self.sp_label = [np.zeros([x]) for x in self.sp_num]
self.trmask = [np.zeros([x]) for x in self.sp_num]
self.temask = [np.ones([x]) for x in self.sp_num]
self.sp_nei = []
self.sp_label_vec = []
self.sp_A = []
self.support = []
self.CalSpMean()
self.CalSpNei()
self.CalSpA()
def ReduceZeros(self,gt):
for i in range(len(self.useful_sp_lab)):
sp_lab=self.useful_sp_lab[i]
n=0
for sp_idx in range(1,sp_lab.max()+1):
sp_pos=np.argwhere(sp_lab==sp_idx)
if gt[sp_pos[:,0],sp_pos[:,1]].sum()==0:
sp_lab[sp_pos[:,0],sp_pos[:,1]]=0
else:
n+=1
sp_lab[sp_pos[:,0],sp_pos[:,1]]=n
self.useful_sp_lab[i]=sp_lab
print('Final number of superpixels is',n,'for',i,'of',len(self.useful_sp_lab))
def CalSpMean(self):
for scale_idx in range(self.scale):
self.gt_tr[self.trpos] = self.gt1d[self.trpos]
mark_mat = np.zeros([self.r*self.c])
mark_mat[self.trpos] = -1
for sp_idx in range(1, self.sp_num[scale_idx]+1): #calculate the sp_mean and sp_label of each super-pixel one-by-one
region_pos_2d = np.argwhere(self.useful_sp_lab[scale_idx] == sp_idx)
region_pos_1d = region_pos_2d[:, 0]*self.c + region_pos_2d[:, 1]
px_num = np.shape(region_pos_2d)[0] #px_num = pixel number in the superpixel
if np.sum(mark_mat[region_pos_1d])<0:#the train_loc in mark_mat is -1
self.trmask[scale_idx][sp_idx-1] = 1
self.temask[scale_idx][sp_idx-1] = 0
region_fea = self.img2d[region_pos_1d, :]
if self.trmask[scale_idx][sp_idx-1] == 1:
region_labels = self.gt_tr[region_pos_1d]
else:
region_labels = self.gt_te[region_pos_1d] # Can you use the gt to calculate the region_labels here?
####
if len(np.delete(np.bincount(region_labels), 0))==0:
print(sp_idx)
####
self.sp_label[scale_idx][sp_idx-1] = np.argmax(np.delete(np.bincount(region_labels), 0))+1 # sp_label is determined by the region labels(the most labels index)
region_pos_idx = np.argwhere(region_labels == self.sp_label[scale_idx][sp_idx-1])
pos1 = region_pos_1d[region_pos_idx]
sp_rps = np.mean(self.img2d[pos1, :], axis = 0) # average of the training pixels in the superpixel
vj = np.sum(np.power(np.matlib.repmat(sp_rps, px_num, 1)-region_fea, 2), axis=1)
vj= np.exp(-1*vj) # coefficient of the pixels(region_fea)
self.sp_mean[scale_idx][sp_idx-1, :] = np.sum(np.reshape(vj, [np.size(vj), 1])*region_fea, axis=0)/np.sum(vj)# weighted average
sp_label_mat = np.zeros([self.sp_num[scale_idx], self.num_classes]) # one-hot coding
for row_idx in range(np.shape(self.sp_label[scale_idx])[0]):
col_idx = int(self.sp_label[scale_idx][row_idx])-1
sp_label_mat[row_idx, col_idx] = 1
self.sp_label_vec.append(self.sp_label[scale_idx])
self.sp_label[scale_idx] = sp_label_mat
def CalSpNei(self): #find the adjacent superpixels of each super-pixel one-by-one
for scale_idx in range(self.scale):
sp_nei=[]
for sp_idx in range(1, self.sp_num[scale_idx]+1):
nei_list = []
region_pos_2d = np.argwhere(self.useful_sp_lab[scale_idx] == sp_idx)
r1 = np.min(region_pos_2d[:, 0])
r2 = np.max(region_pos_2d[:, 0])
c1 = np.min(region_pos_2d[:, 1])
c2 = np.max(region_pos_2d[:, 1])
for r in range(r1, r2+1):#按行遍历,找到最临近的useful_spp_lab
pos1 = np.argwhere(region_pos_2d[:, 0] == r)[:, 0]
try:
min_col = np.min(region_pos_2d[:, 1][pos1])
max_col = np.max(region_pos_2d[:, 1][pos1])
except:
pass #print(region_pos_2d[:, 1][pos1])
nc1 = min_col-1
nc2 = max_col+1
if nc1>=0:
nei_list.append(self.useful_sp_lab[scale_idx][r, nc1])
if nc2<=self.c-1:
nei_list.append(self.useful_sp_lab[scale_idx][r, nc2])
for c in range(c1, c2+1):#按列遍历,找到最临近的useful_spp_lab
pos1 = np.argwhere(region_pos_2d[:, 1] == c)[:, 0]
try:
min_row = np.min(region_pos_2d[:, 0][pos1])
max_row = np.max(region_pos_2d[:, 0][pos1])
except:
pass #print(region_pos_2d[:, 0][pos1])
nr1 = min_row-1
nr2 = max_row+1
if nr1>=0:
nei_list.append(self.useful_sp_lab[scale_idx][nr1, c])
if nr2<=self.r-1:
nei_list.append(self.useful_sp_lab[scale_idx][nr2, c])
nei_list = list(set(nei_list))
nei_list = [int(list_item) for list_item in nei_list]
if 0 in nei_list:
nei_list.remove(0)
sp_nei.append(nei_list if len(nei_list) else [])
self.sp_nei.append(sp_nei)
def CalSpA(self):
for scale_idx in range(self.scale):
sp_A_s1 = np.zeros([self.sp_num[scale_idx], self.sp_num[scale_idx]])
for sp_idx in range(1, self.sp_num[scale_idx]+1):
sp_idx0 = sp_idx-1
cen_sp = self.sp_mean[scale_idx][sp_idx0]
nei_idx = self.sp_nei[scale_idx][sp_idx0] # list
nei_idx0 = np.array([list_item-1 for list_item in nei_idx], dtype=int)
cen_nei = self.sp_mean[scale_idx][nei_idx0, :]
dist1 = self.Eu_dist(cen_sp, cen_nei)
sp_A_s1[sp_idx0, nei_idx0] = dist1
self.sp_A.append(sp_A_s1)
self.sp_A[-1] = self.SymmetrizationMat(self.sp_A[-1])
def AddConnection(self, A):
A1 = A.copy()
num_rows = np.shape(A)[0]
for row_idx in range(num_rows): #two-hops extension the check the rows one-by-one
pos1 = np.argwhere(A[row_idx, :]!=0)
for num_nei1 in range(np.size(pos1)):
nei_ori = A[pos1[num_nei1, 0], :].copy()
pos2 = np.argwhere(nei_ori!=0)[:, 0]
nei1 = self.sp_mean[pos2, :]
dist1 = self.Eu_dist(self.sp_mean[row_idx, :], nei1)
A1[row_idx, pos2] = dist1
A1[row_idx, row_idx] = 0 #set 0 to the diag
return A1
def Eu_dist(self, vec, mat):
rows = np.shape(mat)[0]
mat1 = np.matlib.repmat(vec, rows, 1)
dist1 = np.exp(-0.1*np.sum(np.power(mat1-mat, 2), axis = 1))
return dist1
def SymmetrizationMat(self, mat):
[r, c] = np.shape(mat)
if r!=c:
print('Input is not square matrix')
return
for rows in range(r):
for cols in range(rows, c):
e1 = mat[rows, cols]
e2 = mat[cols, rows]
if e1+e2!=0 and e1*e2 == 0:
mat[rows, cols] = e1+e2
mat[cols, rows] = e1+e2
return mat
def CalSupport(self, A):
num1 = np.shape(A)[0]
A_ = A + 1*np.eye(num1)
D_ = np.sum(A_, 1)
D_05 = np.diag(D_**(-0.5))
support = np.matmul(np.matmul(D_05, A_), D_05)
return support #(D_**(-0.5)A(D_**(-0.5))Z