/
SuperPointNet_gauss2.py
204 lines (171 loc) · 6.67 KB
/
SuperPointNet_gauss2.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
201
"""latest version of SuperpointNet. Use it!
"""
import torch
import torch.nn as nn
from torch.nn.init import xavier_uniform_, zeros_
from models.unet_parts import *
import numpy as np
# from models.SubpixelNet import SubpixelNet
class SuperPointNet_gauss2(torch.nn.Module):
""" Pytorch definition of SuperPoint Network. """
def __init__(self, subpixel_channel=1):
super(SuperPointNet_gauss2, self).__init__()
c1, c2, c3, c4, c5, d1 = 64, 64, 128, 128, 256, 256
det_h = 65
self.inc = inconv(1, c1)
self.down1 = down(c1, c2)
self.down2 = down(c2, c3)
self.down3 = down(c3, c4)
# self.down4 = down(c4, 512)
# self.up1 = up(c4+c3, c2)
# self.up2 = up(c2+c2, c1)
# self.up3 = up(c1+c1, c1)
# self.outc = outconv(c1, subpixel_channel)
self.relu = torch.nn.ReLU(inplace=True)
# self.outc = outconv(64, n_classes)
# Detector Head.
self.convPa = torch.nn.Conv2d(c4, c5, kernel_size=3, stride=1, padding=1)
self.bnPa = nn.BatchNorm2d(c5)
self.convPb = torch.nn.Conv2d(c5, det_h, kernel_size=1, stride=1, padding=0)
self.bnPb = nn.BatchNorm2d(det_h)
# Descriptor Head.
self.convDa = torch.nn.Conv2d(c4, c5, kernel_size=3, stride=1, padding=1)
self.bnDa = nn.BatchNorm2d(c5)
self.convDb = torch.nn.Conv2d(c5, d1, kernel_size=1, stride=1, padding=0)
self.bnDb = nn.BatchNorm2d(d1)
self.output = None
def forward(self, x):
""" Forward pass that jointly computes unprocessed point and descriptor
tensors.
Input
x: Image pytorch tensor shaped N x 1 x patch_size x patch_size.
Output
semi: Output point pytorch tensor shaped N x 65 x H/8 x W/8.
desc: Output descriptor pytorch tensor shaped N x 256 x H/8 x W/8.
"""
# Let's stick to this version: first BN, then relu
x1 = self.inc(x)
x2 = self.down1(x1)
x3 = self.down2(x2)
x4 = self.down3(x3)
# Detector Head.
cPa = self.relu(self.bnPa(self.convPa(x4)))
semi = self.bnPb(self.convPb(cPa))
# Descriptor Head.
cDa = self.relu(self.bnDa(self.convDa(x4)))
desc = self.bnDb(self.convDb(cDa))
dn = torch.norm(desc, p=2, dim=1) # Compute the norm.
desc = desc.div(torch.unsqueeze(dn, 1)) # Divide by norm to normalize.
output = {'semi': semi, 'desc': desc}
self.output = output
return output
def process_output(self, sp_processer):
"""
input:
N: number of points
return: -- type: tensorFloat
pts: tensor [batch, N, 2] (no grad) (x, y)
pts_offset: tensor [batch, N, 2] (grad) (x, y)
pts_desc: tensor [batch, N, 256] (grad)
"""
from utils.utils import flattenDetection
# from models.model_utils import pred_soft_argmax, sample_desc_from_points
output = self.output
semi = output['semi']
desc = output['desc']
# flatten
heatmap = flattenDetection(semi) # [batch_size, 1, H, W]
# nms
heatmap_nms_batch = sp_processer.heatmap_to_nms(heatmap, tensor=True)
# extract offsets
outs = sp_processer.pred_soft_argmax(heatmap_nms_batch, heatmap)
residual = outs['pred']
# extract points
outs = sp_processer.batch_extract_features(desc, heatmap_nms_batch, residual)
# output.update({'heatmap': heatmap, 'heatmap_nms': heatmap_nms, 'descriptors': descriptors})
output.update(outs)
self.output = output
return output
def get_matches(deses_SP):
from models.model_wrap import PointTracker
tracker = PointTracker(max_length=2, nn_thresh=1.2)
f = lambda x: x.cpu().detach().numpy()
# tracker = PointTracker(max_length=2, nn_thresh=1.2)
# print("deses_SP[1]: ", deses_SP[1].shape)
matching_mask = tracker.nn_match_two_way(f(deses_SP[0]).T, f(deses_SP[1]).T, nn_thresh=1.2)
return matching_mask
# print("matching_mask: ", matching_mask.shape)
# f_mask = lambda pts, maks: pts[]
# pts_m = []
# pts_m_res = []
# for i in range(2):
# idx = xs_SP[i][matching_mask[i, :].astype(int), :]
# res = reses_SP[i][matching_mask[i, :].astype(int), :]
# print("idx: ", idx.shape)
# print("res: ", idx.shape)
# pts_m.append(idx)
# pts_m_res.append(res)
# pass
# pts_m = torch.cat((pts_m[0], pts_m[1]), dim=1)
# matches_test = toNumpy(pts_m)
# print("pts_m: ", pts_m.shape)
# pts_m_res = torch.cat((pts_m_res[0], pts_m_res[1]), dim=1)
# # pts_m_res = toNumpy(pts_m_res)
# print("pts_m_res: ", pts_m_res.shape)
# # print("pts_m_res: ", pts_m_res)
# pts_idx_res = torch.cat((pts_m, pts_m_res), dim=1)
# print("pts_idx_res: ", pts_idx_res.shape)
def main():
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = SuperPointNet_gauss2()
model = model.to(device)
# check keras-like model summary using torchsummary
from torchsummary import summary
summary(model, input_size=(1, 240, 320))
## test
image = torch.zeros((2,1,120, 160))
outs = model(image.to(device))
print("outs: ", list(outs))
from utils.print_tool import print_dict_attr
print_dict_attr(outs, 'shape')
from models.model_utils import SuperPointNet_process
params = {
'out_num_points': 500,
'patch_size': 5,
'device': device,
'nms_dist': 4,
'conf_thresh': 0.015
}
sp_processer = SuperPointNet_process(**params)
outs = model.process_output(sp_processer)
print("outs: ", list(outs))
print_dict_attr(outs, 'shape')
# timer
import time
from tqdm import tqdm
iter_max = 50
start = time.time()
print("Start timer!")
for i in tqdm(range(iter_max)):
outs = model(image.to(device))
end = time.time()
print("forward only: ", iter_max/(end - start), " iter/s")
start = time.time()
print("Start timer!")
xs_SP, deses_SP, reses_SP = [], [], []
for i in tqdm(range(iter_max)):
outs = model(image.to(device))
outs = model.process_output(sp_processer)
xs_SP.append(outs['pts_int'].squeeze())
deses_SP.append(outs['pts_desc'].squeeze())
reses_SP.append(outs['pts_offset'].squeeze())
end = time.time()
print("forward + process output: ", iter_max/(end - start), " iter/s")
start = time.time()
print("Start timer!")
for i in tqdm(range(len(xs_SP))):
get_matches([deses_SP[i][0], deses_SP[i][1]])
end = time.time()
print("nn matches: ", iter_max/(end - start), " iters/s")
if __name__ == '__main__':
main()