-
Notifications
You must be signed in to change notification settings - Fork 2
/
dataset_generation.py
291 lines (251 loc) · 12.4 KB
/
dataset_generation.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
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
import tensorflow.compat.v1 as tf
tf.enable_eager_execution()
tf.config.set_visible_devices([], 'GPU')
import argparse
import os
import numpy as np
import pickle
from PIL import Image
from tqdm import tqdm
from waymo_open_dataset import dataset_pb2 as open_dataset
from waymo_open_dataset.utils import camera_segmentation_utils
from dataset_generation_const import *
CAM_TYPE = {open_dataset.CameraName.FRONT_LEFT: 0, open_dataset.CameraName.FRONT: 1, open_dataset.CameraName.FRONT_RIGHT: 2}
def save_train_seq(seq, bc_seq, seq_num, out_dir):
num_cam = len(CAM_TYPE)
temporal_transform = np.zeros((len(seq), 4, 4))
timestamps = np.zeros((len(seq), num_cam))
top_crop_ratio = 60/140
top_offset = RESIZE_DIM[1]*top_crop_ratio
crop_ltrb = (0, top_offset, RESIZE_DIM[0], RESIZE_DIM[1])
final_dim_hw = (RESIZE_DIM[1] - int(top_offset), RESIZE_DIM[0])
rgbs = np.zeros((len(seq), num_cam, final_dim_hw[0], final_dim_hw[1], 3))
extrinsics = np.zeros((len(seq), num_cam, 4, 4))
for (t, frame) in enumerate(seq):
temporal_transform[t] = np.array(frame.pose.transform).reshape((4,4))
for image in frame.images:
if image.name in CAM_TYPE:
f = CAM_TYPE[image.name]
timestamps[t, f] = frame.timestamp_micros
rgb = tf.image.decode_jpeg(image.image).numpy()
rgb_img = Image.fromarray(rgb).resize((RESIZE_DIM)).crop((crop_ltrb))
rgbs[t, f] = np.array(rgb_img).astype('float') / 255
for calibration in frame.context.camera_calibrations:
if calibration.name in CAM_TYPE:
f = CAM_TYPE[calibration.name]
# sensor to vehicle frame
extrinsics[t, f] = np.array(calibration.extrinsic.transform).reshape((4, 4))
origin = np.expand_dims(temporal_transform[0], 0) # 1 x 4 x 4
viewpoint_transform = np.expand_dims(np.linalg.solve(origin, temporal_transform), 1) @ extrinsics # T x C x 4 x 4
viewpoint_transform[:, :, :3, 3] /= DISTANCE_NORMALIZATION_FACTOR
viewpoint_transform = viewpoint_transform[:, :, :-1].reshape((len(seq), num_cam, -1))
timestamps -= timestamps[0]
timestamps /= TIMESTAMP_NORMALIZATION_FACTOR
bc_waypoints = np.zeros((len(bc_seq), 2))
bc_mask = np.zeros((len(bc_seq)))
bc_origin = temporal_transform[-1] # 4 x 4
for (t, frame) in enumerate(bc_seq):
if frame is not None:
bc_mask[t] = 1
bc_transform = np.linalg.solve(bc_origin, np.array(frame.pose.transform).reshape((4,4)))
bc_waypoints[t] = bc_transform[:2, 3].T / DISTANCE_NORMALIZATION_FACTOR # 1 x 2
with open(os.path.join(out_dir, f'{seq_num}.npz'), 'wb') as f:
np.savez_compressed(f, rgb=rgbs,
viewpoint_transform=viewpoint_transform,
time=timestamps,
bc_waypoints=bc_waypoints,
bc_mask=bc_mask)
return
def make_train_seqs(first_seq_num, unique_start_ids, in_dir, out_dir):
seq_num = first_seq_num
train_files = os.path.join(in_dir, '*.tfrecord')
filenames = tf.io.matching_files(train_files)
dataset = tf.data.TFRecordDataset(filenames)
current_seq = []
current_BC = [None for _ in range(BC_LEN)]
current_t = 0
new_seq = True
collect_BC = False
for data in tqdm(dataset):
frame = open_dataset.Frame()
frame.ParseFromString(bytearray(data.numpy()))
new_t = frame.timestamp_micros
# Collecting image frames for autoencoding
if new_seq or (not collect_BC and ((new_t - current_t > STRIDE*1e5 - 0.05e6) and (new_t - current_t < STRIDE*1e5 + 0.05e6))):
if new_seq and new_t in unique_start_ids:
continue
current_seq.append(frame)
new_seq = False
current_t = new_t
# Once we've collected enough image frames, switch to collecting trajectory info
if len(current_seq) == SEQ_LEN:
collect_BC = True
seq_end_t = current_t
# Collecting future trajectory information for BC
elif collect_BC:
elapsed_t = new_t - seq_end_t
index = int(np.round(elapsed_t / (BC_STRIDE*1e5)) - 1)
# Deal with timeskips
if index < 0:
index = np.inf
if index < BC_LEN:
current_BC[index] = frame
if index >= BC_LEN - 1:
save_train_seq(current_seq, current_BC, seq_num, out_dir)
seq_start_t = current_seq[0].timestamp_micros
unique_start_ids[seq_start_t] = seq_num
seq_num += 1
current_seq = []
current_BC = [None for _ in range(BC_LEN)]
new_seq = True
collect_BC = False
current_t = new_t
# The frame we wanted with the correct timestamp is missing, so start over with a new sequence
elif (new_t - current_t < 0) or (new_t - current_t > STRIDE*1e5 + 0.05e6):
if new_t not in unique_start_ids:
current_seq = [frame]
current_t = new_t
new_seq = False
else:
current_seq = []
current_t = new_t
new_seq = True
# Note that if we don't meet either above condition, it just means we haven't reached the next frame with the
# correct timestamp yet
return seq_num
def save_val_seq(seq, seq_num, out_dir):
panoptic_label_inds = range(len(seq))
num_cam = len(CAM_TYPE)
seg_protos = [0 for _ in range(num_cam*SEQ_LEN)]
temporal_transform = np.zeros((len(seq), 4, 4))
timestamps = np.zeros((len(seq), num_cam))
top_crop_ratio = 60/140
top_offset = RESIZE_DIM[1]*top_crop_ratio
crop_ltrb = (0, top_offset, RESIZE_DIM[0], RESIZE_DIM[1])
final_dim_hw = (RESIZE_DIM[1] - int(top_offset), RESIZE_DIM[0])
rgbs = np.zeros((len(seq), num_cam, final_dim_hw[0], final_dim_hw[1], 3))
extrinsics = np.zeros((len(seq), num_cam, 4, 4))
for (t, frame) in enumerate(seq):
temporal_transform[t] = np.array(frame.pose.transform).reshape((4,4))
for image in frame.images:
if image.name in CAM_TYPE:
f = CAM_TYPE[image.name]
timestamps[t, f] = frame.timestamp_micros
rgb = tf.image.decode_jpeg(image.image).numpy()
rgb_img = Image.fromarray(rgb).resize((RESIZE_DIM)).crop((crop_ltrb))
rgbs[t, f] = np.array(rgb_img).astype('float') / 255
if t in panoptic_label_inds:
idx = np.ravel_multi_index((t, f), (SEQ_LEN, num_cam))
seg_protos[idx] = image.camera_segmentation_label
for calibration in frame.context.camera_calibrations:
if calibration.name in CAM_TYPE:
f = CAM_TYPE[calibration.name]
# sensor to vehicle frame
extrinsics[t, f] = np.array(calibration.extrinsic.transform).reshape((4, 4))
origin = np.expand_dims(temporal_transform[0], 0) # 1 x 4 x 4
viewpoint_transform = np.expand_dims(np.linalg.solve(origin, temporal_transform), 1) @ extrinsics # T x C x 4 x 4
viewpoint_transform[:, :, :3, 3] /= DISTANCE_NORMALIZATION_FACTOR
viewpoint_transform = viewpoint_transform[:, :, :-1].reshape((len(seq), num_cam, -1))
timestamps -= timestamps[0]
timestamps /= TIMESTAMP_NORMALIZATION_FACTOR
(panoptic_labels, _, panoptic_label_divisor) = camera_segmentation_utils.decode_multi_frame_panoptic_labels_from_protos(
seg_protos, remap_values=True
)
semantic_segs = np.zeros(rgbs.shape[:-1], dtype='int')
instance_segs = np.zeros(rgbs.shape[:-1], dtype='int')
for (i, label) in enumerate(panoptic_labels):
(semantic_label_front, instance_label_front) = camera_segmentation_utils.decode_semantic_and_instance_labels_from_panoptic_label(
label,
panoptic_label_divisor)
(t, f) = np.unravel_index(i, (SEQ_LEN, num_cam))
semantic_label_img = Image.fromarray(semantic_label_front.astype('uint8').squeeze())
semantic_label_img = semantic_label_img.resize(RESIZE_DIM, resample=Image.Resampling.NEAREST).crop(crop_ltrb)
semantic_segs[t, f] = np.array(semantic_label_img).astype('int')
instance_label_img = Image.fromarray(instance_label_front.astype('uint8').squeeze())
instance_label_img = instance_label_img.resize(RESIZE_DIM, resample=Image.Resampling.NEAREST).crop(crop_ltrb)
instance_segs[t, f] = np.array(instance_label_img).astype('int')
with open(os.path.join(out_dir, f'{seq_num}.npz'), 'wb') as f:
np.savez_compressed(f, rgb=rgbs,
semantic_seg=semantic_segs,
instance_seg=instance_segs,
viewpoint_transform=viewpoint_transform,
time=timestamps)
return
def collate_val_seqs(in_dir):
"""
Find every sequence of frames with panoptic segmentation labels in the dataset.
"""
val_files = os.path.join(in_dir, '*.tfrecord')
filenames = tf.io.matching_files(val_files)
dataset = tf.data.TFRecordDataset(filenames)
seq_dict = {}
for data in tqdm(dataset):
frame = open_dataset.Frame()
frame.ParseFromString(bytearray(data.numpy()))
for image in frame.images:
if image.name in CAM_TYPE:
break
if image.camera_segmentation_label.panoptic_label:
seq_id = image.camera_segmentation_label.sequence_id
if seq_id in seq_dict:
seq_dict[seq_id].append(frame)
else:
seq_dict[seq_id] = [frame]
return seq_dict
def make_val_seqs(out_dir, seq_dict):
"""
Save frame sequences where there are no missing frames.
"""
seq_num = 0
for (_, seq) in tqdm(seq_dict.items()):
seq = sorted(seq, key=lambda frame: frame.timestamp_micros)
new_seq = True
current_seq = []
for frame in seq:
new_t = frame.timestamp_micros
if new_seq or ((new_t - current_t > STRIDE*1e5 - 0.05e6) and (new_t - current_t < STRIDE*1e5 + 0.05e6)):
current_seq.append(frame)
new_seq = False
current_t = new_t
if len(current_seq) == SEQ_LEN:
save_val_seq(current_seq, seq_num, out_dir)
seq_num += 1
current_seq = []
new_seq = True
else:
current_seq = [frame]
current_t = new_t
new_seq = False
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('split', choices=['train', 'val'])
parser.add_argument('--in_dir', default='waymo_open_raw')
parser.add_argument('--out_dir', default='waymo_open')
parser.add_argument('--load_seq_ids', default=None,
help='To resume generating training sequences, load previously generated IDs from file')
parser.add_argument('--save_seq_ids', default=None,
help='To resume generating training sequences later, save newly generated IDs to file')
args = parser.parse_args()
in_dir = os.path.join(args.in_dir, args.split)
out_dir = os.path.join(args.out_dir, args.split)
os.makedirs(out_dir, exist_ok=True)
if args.split == 'train':
if args.load_seq_ids is not None:
with open(args.load_seq_ids, 'rb') as f:
unique_start_ids = pickle.load(f)
else:
unique_start_ids = {}
prev_seq_num = len(unique_start_ids)
new_seq_num = make_train_seqs(prev_seq_num, unique_start_ids, in_dir, out_dir)
if args.save_seq_ids is not None:
with open(args.save_seq_ids, 'wb') as f:
pickle.dump(unique_start_ids, f)
while new_seq_num > prev_seq_num and new_seq_num < MAX_NUM_TRAIN_SEQ:
prev_seq_num = new_seq_num
new_seq_num = make_train_seqs(prev_seq_num, unique_start_ids, in_dir, out_dir)
if args.save_seq_ids is not None:
with open(args.save_seq_ids, 'wb') as f:
pickle.dump(unique_start_ids, f)
elif args.split == 'val':
seq_dict = collate_val_seqs(in_dir)
make_val_seqs(out_dir, seq_dict)