This repository has been archived by the owner on Jan 23, 2023. It is now read-only.
/
heatmap.py
executable file
·241 lines (211 loc) · 8.48 KB
/
heatmap.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
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
#!/usr/bin/env python
import os
import sys
import tqdm
import argparse
import numpy as np
import pandas as pd
from skimage import io
import torch
import torchvision
from osgeo import osr
from osgeo import gdal
sys.path.append('../../model')
import cvig_fov as cvig
Globals = cvig.Globals
device = cvig.device
names = [
'01_rio',
'02_vegas',
'03_paris',
'04_shanghai',
'05_khartoum',
'06_atlanta',
'07_moscow',
'08_mumbai',
'09_san',
'10_dar',
'11_rotterdam',
]
class ImageDataset(torch.utils.data.Dataset):
def __init__(self, paths, transform=None):
self.paths = paths
self.transform = transform
def __len__(self):
return len(self.paths)
def __getitem__(self, idx):
raw = io.imread(self.paths[idx])
image = torch.from_numpy(raw.astype(np.float32).transpose((2, 0, 1)))
data = {'image':image}
if self.transform is not None:
data = self.transform(data)
return data
class TileDataset(torch.utils.data.Dataset):
def __init__(self, source, windows, transform=None):
self.source = source
self.windows = windows
self.transform = transform
def __len__(self):
return len(self.windows)
def __getitem__(self, idx):
mem_path = '/vsimem/tile%s.jpg' % str(idx)
ds = gdal.Translate(mem_path, self.source, projWin=self.windows[idx])
raw = ds.ReadAsArray()
gdal.GetDriverByName('GTiff').Delete(mem_path)
image = torch.from_numpy(raw.astype(np.float32))
data = {'image':image}
if self.transform is not None:
data = self.transform(data)
return data
class ResizeSurface(object):
"""
Resize surface photo to fit model and crop to fov.
"""
def __init__(self, fov=360):
self.fov = fov
self.surface_width = int(self.fov / 360 * Globals.surface_width_max)
def __call__(self, data):
data['image'] = torchvision.transforms.functional.resize(data['image'], (Globals.surface_height_max, self.surface_width))
return data
class ResizeOverhead(object):
"""
Resize overhead image tile to fit model and crop to fov.
"""
def __call__(self, data):
data['image'] = torchvision.transforms.functional.resize(data['image'], (Globals.overhead_size, Globals.overhead_size))
return data
class ImageNormalization(object):
"""
Normalize image values to use with pretrained VGG model
"""
def __init__(self):
self.norm = torchvision.transforms.Normalize(
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]
)
def __call__(self, data):
data['image'] = self.norm(data['image'] / 255.)
return data
class PolarTransform(object):
def __init__(self):
self.transform = cvig.PolarTransform()
def __call__(self, data):
data_renamed = {'overhead':data['image']}
data = self.transform(data_renamed)
return data
def sweep(aoi, bounds, edge, offset, fov, sat_dir, photo_path, csv_path):
# Compute center and window for each satellite tile
center_eastings = []
center_northings = []
windows = []
e2 = edge / 2.
for easting in np.arange(bounds[0] - e2, bounds[2] - e2, offset):
for northing in np.arange(bounds[3] + e2, bounds[1] + e2, -offset):
center_eastings.append(easting + e2)
center_northings.append(northing - e2)
windows.append([easting, northing, easting + edge, northing - edge])
# Load satellite strip
sat_path = os.path.join(sat_dir, names[aoi-1] + '.tif')
sat_file = gdal.Open(sat_path)
# Specify transformations
surface_transform = torchvision.transforms.Compose([
ResizeSurface(fov),
ImageNormalization()
])
overhead_transform = torchvision.transforms.Compose([
ResizeOverhead(),
ImageNormalization(),
PolarTransform()
])
# Load data
surface_set = ImageDataset((photo_path,), surface_transform)
overhead_set = TileDataset(sat_file, windows, overhead_transform)
surface_batch = torch.unsqueeze(surface_set[0]['image'], dim=0).to(device)
overhead_loader = torch.utils.data.DataLoader(overhead_set, batch_size=64, shuffle=False, num_workers=1)
# Load the neural networks
surface_encoder = cvig.FOV_DSM(circ_padding=False).to(device)
overhead_encoder = cvig.FOV_DSM(circ_padding=True).to(device)
surface_encoder.load_state_dict(torch.load('../../model/fov_{}_surface_best.pth'.format(int(fov))))
overhead_encoder.load_state_dict(torch.load('../../model/fov_{}_overhead_best.pth'.format(int(fov))))
surface_encoder.eval()
overhead_encoder.eval()
# Surface photo's features
surface_embed = surface_encoder(surface_batch)
# Overhead images' features
torch.set_grad_enabled(False)
overhead_embed = None
for batch, data in enumerate(tqdm.tqdm(overhead_loader)):
overhead = data['polar'].to(device)
#with torch.set_grad_enabled(False):
overhead_embed_part = overhead_encoder(overhead)
if overhead_embed is None:
overhead_embed = overhead_embed_part
else:
overhead_embed = torch.cat((overhead_embed, overhead_embed_part), dim=0)
# Calculate score for each overhead image
output_width_max = 64
orientation_estimate = cvig.correlation(overhead_embed, surface_embed)
orientations = torch.squeeze(orientation_estimate) * 360 / output_width_max - 180
overhead_cropped_all = cvig.crop_overhead(overhead_embed, orientation_estimate, surface_embed.shape[3])
distances = cvig.l2_distance(overhead_cropped_all, surface_embed)
distances = torch.squeeze(distances)
scores = torch.exp(10. * (1. - distances))
# Save information to disk
df = pd.DataFrame({
'x': center_eastings,
'y': center_northings,
'orientation': orientations.cpu().numpy(),
'dissimilarity': distances.cpu().numpy(),
'score': scores.cpu().numpy()
})
df.to_csv(csv_path, index=False)
def layer(aoi, bounds, sat_dir, layer_path):
sat_path = os.path.join(sat_dir, names[aoi-1] + '.tif')
sat_file = gdal.Open(sat_path)
window = [bounds[0], bounds[3], bounds[2], bounds[1]]
gdal.Translate(layer_path, sat_file, projWin=window)
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('-a', '--aoi',
type=int,
choices=range(1,12),
default=3,
help='SpaceNet AOI of satellite image')
parser.add_argument('-b', '--bounds',
type=float,
nargs=4,
default=(447665.8, 5411329.8, 448184.8, 5411814.8),
metavar=('left', 'bottom', 'right', 'top'),
help='Bounds given as UTM coordinates in this order: min easting, min northing, max easting, max northing')
parser.add_argument('-e', '--edge',
type=float,
default=225,
help='Edge length of satellite imagery tiles [m]')
parser.add_argument('-o', '--offset',
type=float,
default=56.25,
help='Offset between centers of adjacent satellite imagery tiles [m]')
parser.add_argument('-f', '--fov',
type=int,
default=70,
help='Field of view assumed for photo (deg, rounded)')
parser.add_argument('-s', '--satdir',
default='/local_data/geoloc/sat/utm',
help='Folder containing satellite images')
parser.add_argument('-p', '--photopath',
default='img.jpg',
help='Path to surface photo to analyze')
parser.add_argument('-c', '--csvpath',
default='./geomatch.csv',
help='Path to output CSV file path')
parser.add_argument('-l', '--layerpath',
default='./satlayer.tiff',
help='Path to output cropped satellite image')
parser.add_argument('-i', '--image',
action='store_true',
help='Flag to output cropped satellite image')
args = parser.parse_args()
sweep(args.aoi, args.bounds, args.edge, args.offset, args.fov,
args.satdir, args.photopath, args.csvpath)
if args.image:
layer(args.aoi, args.bounds, args.satdir, args.layerpath)