-
Notifications
You must be signed in to change notification settings - Fork 31
/
Copy pathHT.py
executable file
·180 lines (142 loc) · 6.63 KB
/
HT.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
#encoding:utf-8
"""
Deep-Hough-Transform-Line-Priors (ECCV 2020) https://arxiv.org/abs/2007.09493
Yancong Lin, and Silvia Laura Pintea, and Jan C. van Gemert
e-mail: yancong.linATtudelftDOTnl
Vision Lab, Delft University of Technology
MIT license
"""
from torch.nn import functional as F
import math
import numpy as np
import torch
import torch.nn as nn
from scipy import ndimage
import cv2
import sys
import scipy.io as sio
import matplotlib.pyplot as plt
# ####################################HT########################################################
def hough_transform(rows, cols, theta_res, rho_res):
theta = np.linspace(0, 180.0, int(np.ceil(180.0 / theta_res) + 1.0))
theta = theta[0:len(theta) - 1]
### Actually,the offset does not have to be this large, because the origin is located at the image center.
D = np.sqrt((rows - 1) ** 2 + (cols - 1) ** 2)
### replace the line above to reduce unnecessray computation (significantly).
# D = np.sqrt((rows/2) ** 2 + (cols/2) ** 2)
q = np.ceil(D / rho_res)
nrho = 2 * q + 1
rho = np.linspace(-q * rho_res, q * rho_res, int(nrho))
w = np.size(theta)
h = np.size(rho)
cos_value = np.cos(theta * np.pi / 180.0).astype(np.float32)
sin_value = np.sin(theta * np.pi / 180.0).astype(np.float32)
sin_cos = np.concatenate((sin_value[None, :], cos_value[None, :]), axis=0)
### This is much more memory-efficient by shifting the coordinate ####
coords_r, coords_w = np.ones((rows, cols)).nonzero()
coords = np.concatenate((coords_r[:,None], coords_w[:,None]), axis=1).astype(np.float32)
coords[:,0] = rows-coords[:,0]-rows//2
coords[:,1] = coords[:,1] +1 - cols//2
vote_map = (coords @ sin_cos).astype(np.float32)
vote_index = np.zeros((rows * cols, h, w))
for i in range(rows*cols):
for j in range(w):
rhoVal = vote_map[i, j]
rhoIdx = np.nonzero(np.abs(rho - rhoVal) == np.min(np.abs(rho - rhoVal)))[0]
vote_map[i, j] = float(rhoIdx[0])
vote_index[i, rhoIdx[0], j] = 1
### remove all-zero lines in the HT maps ####
vote_rho_idx = vote_index.reshape(rows * cols, h, w).sum(axis=0).sum(axis=1)
vote_index = vote_index[:,vote_rho_idx>0.0 ,:]
### update h, since we remove those HT lines without any votes
### slightly different from the original paper, the HT size in this script is 182x60.
h = (vote_rho_idx>0.0).sum()
return vote_index.reshape(rows, cols, h, w)
# torch.nn.Conv2d(in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1, groups=1, bias=True)
def make_conv_block(in_channels, out_channels, kernel_size=3, stride=1, padding=0, dilation=1, groups=1, bias=False):
layers = []
layers += [nn.Conv2d(in_channels, out_channels, kernel_size, stride, padding, dilation, groups, bias)]
### no batchnorm layers
# layers += [nn.BatchNorm2d(out_channels)]
layers += [nn.ReLU(inplace=True)]
return nn.Sequential(*layers)
class HT(nn.Module):
def __init__(self, vote_index):
super(HT, self).__init__()
self.r, self.c, self.h, self.w = vote_index.size()
self.norm = max(self.r, self.c)
self.vote_index = vote_index.view(self.r * self.c, self.h *self.w)
self.total = vote_index.sum(0).max()
def forward(self, image):
batch, channel, _, _ = image.size()
image = image.view(batch,channel, -1).view(batch*channel, -1)
image = F.relu(image)
HT_map = image @ self.vote_index
### normalization ###
# HT_map = HT_map/self.total
### normalized by max(rows, cols)
HT_map = HT_map/(self.norm)
HT_map = HT_map.view(batch, channel, -1).view(batch, channel, self.h, self.w)
return HT_map
class IHT(nn.Module):
def __init__(self, vote_index):
super(IHT, self).__init__()
self.r, self.c, self.h, self.w = vote_index.size()
self.vote_index = vote_index.view(self.r * self.c, self.h * self.w).t()
def forward(self, input_HT):
batch, channel, _, _ = input_HT.size()
input_HT = F.relu(input_HT)
input_HT = input_HT.view(batch, channel, self.h * self.w).view(batch * channel, self.h * self.w)
IHT_map = input_HT @ self.vote_index
IHT_map = IHT_map.view(batch, channel, self.r*self.c).view(batch, channel, self.r, self.c)
# return IHT_map/float(self.w)
return IHT_map
class HTIHT(nn.Module):
def __init__(self, vote_index, inplanes, outplanes):
super(HTIHT, self).__init__()
self.conv1 = nn.Sequential(*make_conv_block(inplanes, inplanes, kernel_size=(9,1), padding=(4,0), bias=True, groups=inplanes))
self.conv2 = nn.Sequential(*make_conv_block(inplanes, outplanes, kernel_size=(9,1), padding=(4,0), bias=True))
self.conv3 = nn.Sequential(*make_conv_block(outplanes, outplanes, kernel_size=(9,1), padding=(4,0), bias=True))
self.relu = nn.ReLU(inplace=True)
self.tanh = nn.Tanh()
self.ht = HT(vote_index)
self.iht = IHT(vote_index)
filtersize = 4
x = np.zeros(shape=((2 * filtersize + 1)))
x[filtersize] = 1
z = []
for _ in range(0, inplanes):
sigma = np.random.uniform(low=1, high=2.5, size=(1))
y = ndimage.filters.gaussian_filter(x, sigma=sigma, order=2)
y = -y / np.sum(np.abs(y))
z.append(y)
z = np.stack(z)
self.conv1[0].weight.data.copy_(torch.from_numpy(z).unsqueeze(1).unsqueeze(3))
nn.init.kaiming_normal_(self.conv2[0].weight, mode='fan_out', nonlinearity='relu')
nn.init.kaiming_normal_(self.conv3[0].weight, mode='fan_out', nonlinearity='relu')
def forward(self, x, **kwargs):
out = self.ht(x)
out = self.conv1(out)
out = self.conv2(out)
out = self.conv3(out)
out = self.iht(out)
return out
class CAT_HTIHT(nn.Module):
def __init__(self, vote_index, inplanes, outplanes):
super(CAT_HTIHT, self).__init__()
self.htiht = HTIHT(vote_index, inplanes, outplanes)
self.bn = nn.BatchNorm2d(inplanes)
self.relu = nn.ReLU(inplace=True)
self.conv_cat = nn.Conv2d(inplanes+outplanes, inplanes, kernel_size=3, padding=1, bias=False)
def forward(self, x):
x = self.bn(x)
x = self.relu(x)
y = self.htiht(x)
out = self.conv_cat(torch.cat([x,y], dim=1))
return out
if __name__ == "__main__":
### Default settings: (128, 128, 3, 1)
vote_index = hough_transform(rows=128, cols=128, theta_res=3, rho_res=1)
rows, cols, h, w = vote_index.shape
print('vote_index', vote_index.shape)
# sio.savemat('../../vote_index_128_31.mat', {'vote_index': vote_index})