/
preprocess_flyingthings3d_subset.py
206 lines (172 loc) · 7.5 KB
/
preprocess_flyingthings3d_subset.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
import os
import cv2
import shutil
import logging
import argparse
import torch.utils.data
import numpy as np
from tqdm import tqdm
from utils import init_logging, load_fpm, load_flow, disp2pc, save_flow_png
'''
Download the "FlyingThings3D subset" from:
https://lmb.informatik.uni-freiburg.de/resources/datasets/SceneFlowDatasets.en.html
You need to download the following part:
* RGB images (cleanpass)
* Disparity
* Disparity change
* Optical flow
* Flow occlusions
Uncompress them and organize the structure of directory as follows:
/mnt/data/flyingthings3d_subset
├── train
│ ├── disparity
│ ├── disparity_change
│ ├── disparity_occlusions
│ ├── flow
│ ├── flow_occlusions
│ └── image_clean
└── val
├── disparity
├── disparity_change
├── disparity_occlusions
├── flow
├── flow_occlusions
└── image_clean
Then preprocess the data:
python preprocess_flyingthings3d_subset.py --input_dir /mnt/data/flyingthings3d_subset
'''
parser = argparse.ArgumentParser()
parser.add_argument('--input_dir', required=True, help='Path to the FlyingThings3D subset')
parser.add_argument('--output_dir', required=False, default='datasets/flyingthings3d_subset_prep')
parser.add_argument('--n_points', required=False, default=[32768, 8192])
parser.add_argument('--max_depth', required=False, default=35.0)
parser.add_argument('--remove_occluded_points', action='store_true')
args = parser.parse_args()
class Preprocessor(torch.utils.data.Dataset):
def __init__(self, input_dir, output_dir, split, n_points, max_depth, remove_occluded_points):
super(Preprocessor, self).__init__()
self.input_dir = input_dir
self.output_dir = output_dir
self.split = split
self.n_points = n_points
self.max_depth = max_depth
self.remove_occluded_points = remove_occluded_points
self.indices = []
for filename in os.listdir(os.path.join(input_dir, split, 'flow', 'left', 'into_future')):
index = int(filename.split('.')[0])
self.indices.append(index)
def __len__(self):
return len(self.indices)
def __getitem__(self, i):
np.random.seed(0)
index1 = self.indices[i]
index2 = index1 + 1
# camera intrinsics
baseline, f, cx, cy = 1.0, 1050.0, 479.5, 269.5
# load data
disp1 = -load_fpm(os.path.join(
self.input_dir, self.split, 'disparity', 'left', '%07d.pfm' % index1
))
disp2 = -load_fpm(os.path.join(
self.input_dir, self.split, 'disparity', 'left', '%07d.pfm' % index2
))
disp1_change = -load_fpm(os.path.join(
self.input_dir, self.split, 'disparity_change', 'left', 'into_future', '%07d.pfm' % index1
))
flow_2d = load_flow(os.path.join(
self.input_dir, self.split, 'flow', 'left', 'into_future', '%07d.flo' % index1
))
occ_mask_2d = cv2.imread(os.path.join(
self.input_dir, self.split, 'flow_occlusions', 'left', 'into_future', '%07d.png' % index1
))
occ_mask_2d = occ_mask_2d[..., 0] > 1
if self.remove_occluded_points:
pc1 = disp2pc(disp1, baseline, f, cx, cy)
pc2 = disp2pc(disp1 + disp1_change, baseline, f, cx, cy, flow_2d)
# apply non-occlusion mask
noc_mask_2d = np.logical_not(occ_mask_2d)
pc1, pc2 = pc1[noc_mask_2d], pc2[noc_mask_2d]
# apply depth mask
mask = np.logical_and(pc1[..., -1] < self.max_depth, pc2[..., -1] < self.max_depth)
pc1, pc2 = pc1[mask], pc2[mask]
# NaN check
mask = np.logical_not(np.isnan(np.sum(pc1, axis=-1) + np.sum(pc2, axis=-1)))
pc1, pc2 = pc1[mask], pc2[mask]
# compute scene flow
flow_3d = pc2 - pc1
occ_mask_3d = np.zeros(len(pc1), dtype=np.bool)
else:
pc1 = disp2pc(disp1, baseline, f, cx, cy)
pc2 = disp2pc(disp2, baseline, f, cx, cy)
flow_3d = disp2pc(disp1 + disp1_change, baseline, f, cx, cy, flow_2d) - pc1
# apply depth mask
mask1 = (pc1[..., -1] < self.max_depth)
mask2 = (pc2[..., -1] < self.max_depth)
pc1, pc2, flow_3d, occ_mask_3d = pc1[mask1], pc2[mask2], flow_3d[mask1], occ_mask_2d[mask1]
# NaN check
mask1 = np.logical_not(np.isnan(np.sum(pc1, axis=-1) + np.sum(flow_3d, axis=-1)))
mask2 = np.logical_not(np.isnan(np.sum(pc2, axis=-1)))
pc1, pc2, flow_3d, occ_mask_3d = pc1[mask1], pc2[mask2], flow_3d[mask1], occ_mask_3d[mask1]
# random sampling
indices1 = np.random.choice(pc1.shape[0], size=min(self.n_points, pc1.shape[0]), replace=False)
indices2 = np.random.choice(pc2.shape[0], size=min(self.n_points, pc2.shape[0]), replace=False)
pc1, pc2, flow_3d, occ_mask_3d = pc1[indices1], pc2[indices2], flow_3d[indices1], occ_mask_3d[indices1]
# save point clouds and occ mask
np.savez(
os.path.join(self.output_dir, self.split, 'pc', '%07d.npz' % index1),
pc1=pc1, pc2=pc2
)
np.save(
os.path.join(self.output_dir, self.split, 'occ_mask_3d', '%07d.npy' % index1),
np.packbits(occ_mask_3d)
)
# mask regions moving extremely fast
flow_mask = np.logical_and(np.abs(flow_2d[..., 0]) < 500, np.abs(flow_2d[..., 1]) < 500)
flow_2d[np.logical_not(flow_mask)] = 0.0
# save ground-truth flow
save_flow_png(
os.path.join(self.output_dir, self.split, 'flow_2d', '%07d.png' % index1),
flow_2d, flow_mask
)
np.save(
os.path.join(self.output_dir, self.split, 'flow_3d', '%07d.npy' % index1),
flow_3d
)
return 0
def main():
for split_idx, split in enumerate(['train', 'val']):
if not os.path.exists(os.path.join(args.input_dir, split)):
continue
logging.info('Processing "%s" split...' % split)
os.makedirs(os.path.join(args.output_dir, split, 'pc'), exist_ok=True)
os.makedirs(os.path.join(args.output_dir, split, 'flow_2d'), exist_ok=True)
os.makedirs(os.path.join(args.output_dir, split, 'flow_3d'), exist_ok=True)
os.makedirs(os.path.join(args.output_dir, split, 'occ_mask_3d'), exist_ok=True)
if not os.path.exists(os.path.join(args.output_dir, split, 'image')):
logging.info('Copying images...')
shutil.copytree(
src=os.path.join(args.input_dir, split, 'image_clean', 'left'),
dst=os.path.join(args.output_dir, split, 'image_clean')
)
if not os.path.exists(os.path.join(args.output_dir, split, 'occ_mask_2d')):
logging.info('Copying occ_mask_2d...')
shutil.copytree(
src=os.path.join(args.input_dir, split, 'flow_occlusions', 'left', 'into_future'),
dst=os.path.join(args.output_dir, split, 'occ_mask_2d')
)
logging.info('Generating point clouds...')
preprocessor = Preprocessor(
args.input_dir,
args.output_dir,
split,
args.n_points[split_idx],
args.max_depth,
args.remove_occluded_points,
)
preprocessor = torch.utils.data.DataLoader(dataset=preprocessor, num_workers=4)
for _ in tqdm(preprocessor):
pass
if __name__ == '__main__':
init_logging()
main()
logging.info('All done.')