Datasets:

License:
sephyli commited on
Commit
5a79da9
·
1 Parent(s): ec992e7

Delete openscene-v1.1/process_data

Browse files
openscene-v1.1/process_data/__init__.py DELETED
File without changes
openscene-v1.1/process_data/create_nuplan_data_with_vis.py DELETED
@@ -1,446 +0,0 @@
1
- import argparse
2
- import shutil
3
- from typing import Dict, List
4
-
5
- # import mmcv
6
- import numpy as np
7
- from os import listdir
8
- from os.path import isfile, join
9
-
10
- from pyquaternion import Quaternion
11
-
12
- import cv2
13
-
14
- from tqdm import tqdm
15
-
16
- import os
17
-
18
- import multiprocessing
19
- import pickle
20
- from nuplan.common.actor_state.state_representation import StateSE2
21
- from nuplan.common.maps.abstract_map import AbstractMap
22
- from nuplan.common.maps.nuplan_map.map_factory import get_maps_api
23
-
24
- from nuplan.database.nuplan_db_orm.nuplandb_wrapper import NuPlanDBWrapper
25
- from nuplan.database.nuplan_db_orm.lidar import Lidar
26
- from nuplan.database.nuplan_db.nuplan_scenario_queries import (
27
- get_traffic_light_status_for_lidarpc_token_from_db,
28
- get_images_from_lidar_tokens,
29
- get_cameras,
30
- )
31
- from nuplan.planning.scenario_builder.nuplan_db.nuplan_scenario import CameraChannel
32
- from navsim.common.extraction.driving_command import get_driving_command
33
-
34
- from helpers.multiprocess_helper import get_scenes_per_thread
35
- from helpers.canbus import CanBus
36
- from helpers.multisweep_helper import obtain_sensor2top
37
-
38
- NUPLAN_MAPS_ROOT = os.environ["NUPLAN_MAPS_ROOT"]
39
- filtered_classes = ["traffic_cone", "barrier", "czone_sign", "generic_object"]
40
-
41
-
42
- def create_nuplan_info(
43
- nuplan_db_wrapper: NuPlanDBWrapper, db_names: List[str], args
44
- ):
45
- nuplan_sensor_root = args.nuplan_sensor_path
46
-
47
- # get all db files & assign db files for current thread.
48
-
49
-
50
- scene_dict = {}
51
- log_sensors = os.listdir(nuplan_sensor_root)
52
-
53
- # For each sequence...
54
- for log_db_name in db_names:
55
- log_db = nuplan_db_wrapper.get_log_db(log_db_name)
56
- log_name = log_db.log_name
57
- log_token = log_db.log.token
58
- map_location = log_db.log.location
59
- vehicle_name = log_db.log.vehicle_name
60
-
61
- # NOTE: I am unsure why "us-nv-las-vegas-strip" is saved as "las_vegas" in db logs.
62
- map_name = map_location if map_location != "las_vegas" else "us-nv-las-vegas-strip"
63
- map_api = get_maps_api(NUPLAN_MAPS_ROOT, "nuplan-maps-v1.0", map_name) # NOTE: lru cached
64
-
65
- log_file = os.path.join(nuplan_db_path, log_db_name + ".db")
66
- if log_db_name not in log_sensors:
67
- continue
68
-
69
- frame_idx = 0
70
-
71
- # list (sequence) of point clouds (each frame).
72
- lidar_pc_list = log_db.lidar_pc
73
- lidar_pcs = lidar_pc_list
74
-
75
- # get log cam infos
76
- log_cam_infos = {}
77
- for cam in get_cameras(log_file, [str(channel.value) for channel in CameraChannel]):
78
- intrinsics = np.array(pickle.loads(cam.intrinsic), dtype=np.float32)
79
- translation = np.array(pickle.loads(cam.translation), dtype=np.float32)
80
- rotation = np.array(pickle.loads(cam.rotation), dtype=np.float32)
81
- rotation = Quaternion(rotation).rotation_matrix
82
- distortion = np.array(pickle.loads(cam.distortion), dtype=np.float32)
83
- c = dict(
84
- intrinsic=intrinsics,
85
- distortion=distortion,
86
- translation=translation,
87
- rotation=rotation,
88
- )
89
- log_cam_infos[cam.token] = c
90
-
91
- # Find the first valid point clouds, with all 8 cameras available.
92
- for start_idx in range(0, len(lidar_pcs)):
93
- retrieved_images = get_images_from_lidar_tokens(
94
- log_file,
95
- [lidar_pcs[start_idx].token],
96
- [str(channel.value) for channel in CameraChannel],
97
- )
98
- if len(list(retrieved_images)) == 8:
99
- break
100
-
101
- # Find the true LiDAR start_idx with the minimum timestamp difference with CAM_F0.
102
- retrieved_images_0 = get_images_from_lidar_tokens(
103
- log_file, [lidar_pcs[start_idx].token], ["CAM_F0"]
104
- )
105
- diff_0 = abs(list(retrieved_images_0)[0].timestamp - lidar_pcs[start_idx].timestamp)
106
- retrieved_images_1 = get_images_from_lidar_tokens(
107
- log_file, [lidar_pcs[start_idx + 1].token], ["CAM_F0"]
108
- )
109
- diff_1 = abs(list(retrieved_images_1)[0].timestamp - lidar_pcs[start_idx + 1].timestamp)
110
- start_idx = start_idx if diff_0 < diff_1 else start_idx + 1
111
-
112
- # Find key_frames (controled by args.sample_interval)
113
- lidar_pc_list = lidar_pc_list[start_idx :: args.sample_interval]
114
- index = -1
115
- for lidar_pc in tqdm(lidar_pc_list, dynamic_ncols=True):
116
- index += 1
117
-
118
- # LiDAR attributes.
119
- lidar_pc_token = lidar_pc.token
120
- scene_token = lidar_pc.scene_token
121
- pc_file_name = lidar_pc.filename
122
- next_token = lidar_pc.next_token
123
- prev_token = lidar_pc.prev_token
124
- lidar_token = lidar_pc.lidar_token
125
- time_stamp = lidar_pc.timestamp
126
- scene_name = "log-" + lidar_pc.scene.name
127
- lidar_boxes = lidar_pc.lidar_boxes
128
- roadblock_ids = [
129
- str(roadblock_id)
130
- for roadblock_id in str(lidar_pc.scene.roadblock_ids).split(" ")
131
- if len(roadblock_id) > 0
132
- ]
133
-
134
- # Saving configurations.
135
- if scene_token not in scene_dict.keys():
136
- scene_dict[scene_token] = []
137
- frame_idx = 0
138
- if frame_idx == 0:
139
- scene_dict[scene_token] = []
140
-
141
- can_bus = CanBus(lidar_pc).tensor
142
- lidar = log_db.session.query(Lidar).filter(Lidar.token == lidar_token).all()
143
- pc_file_path = os.path.join(args.nuplan_sensor_path, pc_file_name)
144
-
145
- if not os.path.exists(pc_file_path): # some lidar files are missing.
146
- # print(pc_file_path)
147
- with open("./nofile.log", "a") as f:
148
- f.write(pc_file_path)
149
- f.write("\n")
150
- continue
151
-
152
- traffic_lights = []
153
- for traffic_light_status in get_traffic_light_status_for_lidarpc_token_from_db(
154
- log_file, lidar_pc_token
155
- ):
156
- lane_connector_id: int = traffic_light_status.lane_connector_id
157
- is_red: bool = traffic_light_status.status.value == 2
158
- traffic_lights.append((lane_connector_id, is_red))
159
-
160
- ego_pose = StateSE2(
161
- lidar_pc.ego_pose.x,
162
- lidar_pc.ego_pose.y,
163
- lidar_pc.ego_pose.quaternion.yaw_pitch_roll[0],
164
- )
165
- driving_command = get_driving_command(ego_pose, map_api, roadblock_ids)
166
-
167
- info = {
168
- "token": lidar_pc_token,
169
- "frame_idx": frame_idx,
170
- "timestamp": time_stamp,
171
- "log_name": log_name,
172
- "log_token": log_token,
173
- "scene_name": scene_name,
174
- "scene_token": scene_token,
175
- "map_location": map_location,
176
- "roadblock_ids": roadblock_ids,
177
- "vehicle_name": vehicle_name,
178
- "can_bus": can_bus,
179
- "lidar_path": pc_file_name, # use the relative path.
180
- "lidar2ego_translation": lidar[0].translation_np,
181
- "lidar2ego_rotation": [
182
- lidar[0].rotation.w,
183
- lidar[0].rotation.x,
184
- lidar[0].rotation.y,
185
- lidar[0].rotation.z,
186
- ],
187
- "ego2global_translation": can_bus[:3],
188
- "ego2global_rotation": can_bus[3:7],
189
- "ego_dynamic_state": [
190
- lidar_pc.ego_pose.vx,
191
- lidar_pc.ego_pose.vy,
192
- lidar_pc.ego_pose.acceleration_x,
193
- lidar_pc.ego_pose.acceleration_y,
194
- ],
195
- "traffic_lights": traffic_lights,
196
- "driving_command": driving_command,
197
- "cams": dict(),
198
- "prev_sweep_token": prev_token,
199
- "next_sweep_token": next_token,
200
- "sweeps": [],
201
- }
202
- info["sample_prev"] = None
203
- info["sample_next"] = None
204
- if index > 0: # find prev.
205
- info["sample_prev"] = lidar_pc_list[index - 1].token
206
- if index < len(lidar_pc_list) - 1: # find next.
207
- next_key_token = lidar_pc_list[index + 1].token
208
- next_key_scene = lidar_pc_list[index + 1].scene_token
209
- info["sample_next"] = next_key_token
210
- else:
211
- next_key_token, next_key_scene = None, None
212
-
213
- if next_key_token == None or next_key_token == "":
214
- frame_idx = 0
215
- else:
216
- if next_key_scene != scene_token:
217
- frame_idx = 0
218
- else:
219
- frame_idx += 1
220
-
221
- # Parse lidar2ego translation.
222
- l2e_r = info["lidar2ego_rotation"]
223
- l2e_t = info["lidar2ego_translation"]
224
- e2g_r = info["ego2global_rotation"]
225
- e2g_t = info["ego2global_translation"]
226
- l2e_r_mat = Quaternion(l2e_r).rotation_matrix
227
- e2g_r_mat = Quaternion(e2g_r).rotation_matrix
228
-
229
- # add lidar2global: map point coord in lidar to point coord in the global
230
- l2e = np.eye(4)
231
- l2e[:3, :3] = l2e_r_mat
232
- l2e[:3, -1] = l2e_t
233
- e2g = np.eye(4)
234
- e2g[:3, :3] = e2g_r_mat
235
- e2g[:3, -1] = e2g_t
236
- lidar2global = np.dot(e2g, l2e)
237
- info["ego2global"] = e2g
238
- info["lidar2ego"] = l2e
239
- info["lidar2global"] = lidar2global
240
-
241
- # obtain 8 image's information per frame
242
- retrieved_images = get_images_from_lidar_tokens(
243
- log_file, [lidar_pc.token], [str(channel.value) for channel in CameraChannel]
244
- )
245
- cams = {}
246
- for img in retrieved_images:
247
- channel = img.channel
248
- filename = img.filename_jpg
249
- filepath = os.path.join(args.nuplan_sensor_path, filename)
250
- if not os.path.exists(filepath):
251
- frame_str = f"{log_db_name}, {lidar_pc_token}"
252
- tqdm.tqdm.write(f"camera file missing: {frame_str}")
253
- continue
254
- cam_info = log_cam_infos[img.camera_token]
255
- cams[channel] = dict(
256
- data_path=filename, # use the relative path.
257
- sensor2lidar_rotation=cam_info["rotation"],
258
- sensor2lidar_translation=cam_info["translation"],
259
- cam_intrinsic=cam_info["intrinsic"],
260
- distortion=cam_info["distortion"],
261
- )
262
- if len(cams) != 8:
263
- frame_str = f"{log_db_name}, {lidar_pc_token}"
264
- tqdm.write(f"not all cameras are available: {frame_str}")
265
- continue
266
- info["cams"] = cams
267
-
268
- # parse sweeps if assigned.
269
- sweeps = []
270
- tmp_info = info
271
- count = 0
272
- while len(sweeps) < args.max_sweeps:
273
- if tmp_info["prev_sweep_token"] == None:
274
- break
275
-
276
- # Get the previous sweep and update info to previous sweep.
277
- sweep = obtain_sensor2top(
278
- tmp_info["prev_sweep_token"], log_db, l2e_t, l2e_r_mat, e2g_t, e2g_r_mat, args
279
- )
280
-
281
- # Save sweeps in every sweep_interval.
282
- tmp_info = sweep
283
- if count == args.sweep_interval:
284
- if os.path.exists(sweep["data_path"]):
285
- sweeps.append(sweep)
286
- count = 0
287
- else:
288
- count += 1
289
- info["sweeps"] = sweeps
290
-
291
- # Parse 3D object labels.
292
- if not args.is_test:
293
- if args.filter_instance:
294
- fg_lidar_boxes = [
295
- box for box in lidar_boxes if box.category.name not in filtered_classes
296
- ]
297
- else:
298
- fg_lidar_boxes = lidar_boxes
299
-
300
- instance_tokens = [item.token for item in fg_lidar_boxes]
301
- track_tokens = [item.track_token for item in fg_lidar_boxes]
302
-
303
- inv_ego_r = lidar_pc.ego_pose.trans_matrix_inv
304
- ego_yaw = lidar_pc.ego_pose.quaternion.yaw_pitch_roll[0]
305
-
306
- locs = np.array(
307
- [
308
- np.dot(
309
- inv_ego_r[:3, :3],
310
- (b.translation_np - lidar_pc.ego_pose.translation_np).T,
311
- ).T
312
- for b in fg_lidar_boxes
313
- ]
314
- ).reshape(-1, 3)
315
- dims = np.array([[b.length, b.width, b.height] for b in fg_lidar_boxes]).reshape(
316
- -1, 3
317
- )
318
- rots = np.array([b.yaw for b in fg_lidar_boxes]).reshape(-1, 1)
319
- rots = rots - ego_yaw
320
-
321
- velocity = np.array([[b.vx, b.vy] for b in fg_lidar_boxes]).reshape(-1, 2)
322
- velocity_3d = np.array([[b.vx, b.vy, b.vz] for b in fg_lidar_boxes]).reshape(-1, 3)
323
-
324
- # convert velo from global to lidar: only need the rotation matrix
325
- for i in range(len(fg_lidar_boxes)):
326
- velo = np.array([*velocity[i], 0.0])
327
- velo = velo @ np.linalg.inv(e2g_r_mat).T @ np.linalg.inv(l2e_r_mat).T
328
- velocity[i] = velo[:2]
329
-
330
- for i in range(len(fg_lidar_boxes)):
331
- velo = velocity_3d[i]
332
- velo = velo @ np.linalg.inv(e2g_r_mat).T @ np.linalg.inv(l2e_r_mat).T
333
- velocity_3d[i] = velo
334
-
335
- names = [box.category.name for box in fg_lidar_boxes]
336
- names = np.array(names)
337
- gt_boxes_nuplan = np.concatenate([locs, dims, rots], axis=1)
338
- info["anns"] = dict(
339
- gt_boxes=gt_boxes_nuplan,
340
- gt_names=names,
341
- gt_velocity_3d=velocity_3d.reshape(-1, 3),
342
- instance_tokens=instance_tokens,
343
- track_tokens=track_tokens,
344
- )
345
- scene_dict[scene_token].append(info)
346
- del map_api
347
-
348
- pkl_file_path = f"{args.out_dir}/{log_name}.pkl"
349
- os.makedirs(args.out_dir, exist_ok=True)
350
- with open(pkl_file_path, "wb") as f:
351
- pickle.dump(dict(scene_dict), f, protocol=pickle.HIGHEST_PROTOCOL)
352
-
353
-
354
- def parse_args():
355
- parser = argparse.ArgumentParser(description="Train a detector")
356
- parser.add_argument(
357
- "--thread-num", type=int, default=50, help="number of threads for multi-processing."
358
- )
359
-
360
- # directory configurations.
361
- parser.add_argument("--nuplan-root-path", help="the path to nuplan root path.")
362
- parser.add_argument("--nuplan-db-path", help="the dir saving nuplan db.")
363
- parser.add_argument("--nuplan-sensor-path", help="the dir to nuplan sensor data.")
364
- parser.add_argument("--nuplan-map-version", help="nuplan mapping dataset version.")
365
- parser.add_argument("--nuplan-map-root", help="path to nuplan map data.")
366
- parser.add_argument("--out-dir", help="output path.")
367
-
368
- # data configurations.
369
- parser.add_argument("--max-sweeps", type=int, default=10, help="number of point cloud sweeps.")
370
- parser.add_argument(
371
- "--sweep-interval", type=int, default=5, help="interval of point cloud sweeps."
372
- )
373
- parser.add_argument(
374
- "--sample-interval", type=int, default=10, help="interval of key frame samples."
375
- )
376
- parser.add_argument(
377
- "--scene-process-type",
378
- type=str,
379
- default="skip",
380
- help="process type when a scene is processed.",
381
- )
382
-
383
- # TODO.
384
- parser.add_argument("--save-bev-images", action="store_true", help="XXX")
385
- parser.add_argument("--save_surround_images", action="store_true", help="XXX")
386
-
387
- # split.
388
- parser.add_argument("--is-test", action="store_true", help="Dealing with Test set data.")
389
- parser.add_argument(
390
- "--filter-instance", action="store_true", help="Ignore instances in filtered_classes."
391
- )
392
- parser.add_argument("--split", type=str, default="train", help="Train/Val/Test set.")
393
-
394
- args = parser.parse_args()
395
- return args
396
-
397
-
398
- if __name__ == "__main__":
399
- args = parse_args()
400
-
401
- nuplan_root_path = args.nuplan_root_path
402
- nuplan_db_path = args.nuplan_db_path
403
- nuplan_sensor_path = args.nuplan_sensor_path
404
- nuplan_map_version = args.nuplan_map_version
405
- nuplan_map_root = args.nuplan_map_root
406
- out_dir = args.out_dir
407
-
408
- nuplan_db_wrapper = NuPlanDBWrapper(
409
- data_root=nuplan_root_path,
410
- map_root=nuplan_map_root,
411
- db_files=nuplan_db_path,
412
- map_version=nuplan_map_version,
413
- )
414
-
415
- nuplan_db_path = args.nuplan_db_path
416
- db_names_with_extension = [
417
- f for f in listdir(nuplan_db_path) if isfile(join(nuplan_db_path, f))
418
- ]
419
- db_names = [name[:-3] for name in db_names_with_extension]
420
- db_names.sort()
421
-
422
- print(db_names)
423
-
424
- db_names_split = np.split(np.array(db_names), args.thread_num)
425
-
426
-
427
- manager = multiprocessing.Manager()
428
- # return_dict = manager.dict()
429
- threads = []
430
- for x in range(args.thread_num):
431
- t = multiprocessing.Process(
432
- target=create_nuplan_info,
433
- name=str(x),
434
- args=(nuplan_db_wrapper, db_names_split[x], args),
435
- )
436
- threads.append(t)
437
- for thr in threads:
438
- thr.start()
439
- for thr in threads:
440
- if thr.is_alive():
441
- thr.join()
442
-
443
- # pkl_file_path = f"{args.out_dir}/{args.split}.pkl"
444
- # os.makedirs(args.out_dir, exist_ok=True)
445
- # with open(pkl_file_path, "wb") as f:
446
- # pickle.dump(dict(return_dict), f, protocol=pickle.HIGHEST_PROTOCOL)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
openscene-v1.1/process_data/driving_command.py DELETED
@@ -1,178 +0,0 @@
1
- from re import A
2
- from typing import Dict, List, Tuple
3
- import warnings
4
-
5
- import numpy as np
6
- import numpy.typing as npt
7
-
8
- from shapely.geometry import Point
9
-
10
- from nuplan.common.maps.abstract_map import AbstractMap
11
- from nuplan.common.actor_state.state_representation import StateSE2
12
- from nuplan.common.maps.maps_datatypes import SemanticMapLayer
13
- from nuplan.common.maps.abstract_map_objects import (
14
- LaneGraphEdgeMapObject,
15
- RoadBlockGraphEdgeMapObject,
16
- )
17
-
18
- from navsim.planning.simulation.planner.pdm_planner.utils.route_utils import (
19
- route_roadblock_correction,
20
- )
21
- from navsim.planning.simulation.planner.pdm_planner.utils.graph_search.dijkstra import (
22
- Dijkstra,
23
- )
24
- from navsim.planning.simulation.planner.pdm_planner.utils.pdm_path import PDMPath
25
- from navsim.planning.simulation.planner.pdm_planner.utils.pdm_geometry_utils import (
26
- convert_absolute_to_relative_se2_array,
27
- )
28
- from navsim.planning.simulation.planner.pdm_planner.utils.pdm_enums import (
29
- SE2Index,
30
- )
31
-
32
- # shapely runtime warning
33
- warnings.filterwarnings("ignore", category=RuntimeWarning)
34
-
35
-
36
- def get_driving_command(
37
- ego_pose: StateSE2,
38
- map_api: AbstractMap,
39
- route_roadblock_ids: List[str],
40
- distance: float = 20,
41
- lateral_offset: float = 2,
42
- ) -> npt.NDArray[np.int]:
43
- """
44
- Creates the one-hot (left, forward, right)driving command for the ego vehicle
45
- :param ego_pose: (x, y, heading) object for global ego pose
46
- :param map_api: nuPlan's map interface
47
- :param route_roadblock_ids: roadblock ids of route
48
- :param distance: longitudinal distance to interpolate on th centerline, defaults to 10
49
- :param lateral_offset: lateral offset for left/right threshold, to defaults to 2
50
- :return: numpy one-hot array of driving_command
51
- """
52
-
53
- driving_command = np.zeros((3,), dtype=int) # one-hot: (left, forward, right)
54
-
55
- # If no route available, go straight
56
- if len(route_roadblock_ids) == 0:
57
- driving_command[1] = 1
58
- return driving_command
59
-
60
- # Apply route correction on route_roadblock_ids
61
- route_roadblock_dict, _ = get_route_dicts(map_api, route_roadblock_ids)
62
- corrected_route_roadblock_ids = route_roadblock_correction(
63
- ego_pose, map_api, route_roadblock_dict
64
- )
65
- route_roadblock_dict, route_lane_dict = get_route_dicts(map_api, corrected_route_roadblock_ids)
66
-
67
- # Find the nearest lane, graph search, and centerline extraction
68
- current_lane = get_current_lane(ego_pose, route_lane_dict)
69
- discrete_centerline = get_discrete_centerline(
70
- current_lane, route_roadblock_dict, route_lane_dict
71
- )
72
- centerline = PDMPath(discrete_centerline)
73
-
74
- # Interpolate target distance on centerline
75
- current_progress = centerline.project(Point(*ego_pose.array))
76
- target_progress = current_progress + distance
77
-
78
- current_pose_array, target_pose_array = centerline.interpolate(
79
- [current_progress, target_progress], as_array=True
80
- )
81
- target_pose_array = convert_absolute_to_relative_se2_array(
82
- StateSE2(*current_pose_array), target_pose_array[None,...]
83
- )[0]
84
-
85
- # Threshold for driving command
86
- if target_pose_array[SE2Index.Y] >= lateral_offset:
87
- driving_command[0] = 1
88
- elif target_pose_array[SE2Index.Y] <= -lateral_offset:
89
- driving_command[2] = 1
90
- else:
91
- driving_command[1] = 1
92
-
93
- # delete some variables for memory management
94
- del route_roadblock_dict, route_lane_dict, _, centerline
95
- return driving_command
96
-
97
-
98
- def get_route_dicts(
99
- map_api: AbstractMap, route_roadblock_ids: List[str]
100
- ) -> Tuple[Dict[str, RoadBlockGraphEdgeMapObject], Dict[str, LaneGraphEdgeMapObject]]:
101
- """
102
- Loads the roadblock and lane dicts
103
- :param map_api: nuPlan's map interface
104
- :param route_roadblock_ids: roadblock ids of route
105
- :return: tuple of roadblock and lane dicts
106
- """
107
-
108
- # remove repeated ids while remaining order in list
109
- route_roadblock_ids = list(dict.fromkeys(route_roadblock_ids))
110
-
111
- route_roadblock_dict: Dict[str, RoadBlockGraphEdgeMapObject] = {}
112
- route_lane_dict: Dict[str, LaneGraphEdgeMapObject] = {}
113
-
114
- for id_ in route_roadblock_ids:
115
- block = map_api.get_map_object(id_, SemanticMapLayer.ROADBLOCK)
116
- block = block or map_api.get_map_object(id_, SemanticMapLayer.ROADBLOCK_CONNECTOR)
117
- route_roadblock_dict[block.id] = block
118
- for lane in block.interior_edges:
119
- route_lane_dict[lane.id] = lane
120
-
121
- return route_roadblock_dict, route_lane_dict
122
-
123
-
124
- def get_current_lane(
125
- ego_pose: StateSE2, route_lane_dict: Dict[str, LaneGraphEdgeMapObject]
126
- ) -> LaneGraphEdgeMapObject:
127
- """
128
- Find current lane, either if intersection with ego pose, or by distance.
129
- :param ego_pose: (x, y, heading) object for global ego pose
130
- :param route_lane_dict: Dictionary of roadblock ids and objects
131
- :return: Lane object
132
- """
133
-
134
- closest_distance = np.inf
135
- starting_lane = None
136
- for edge in route_lane_dict.values():
137
- if edge.contains_point(ego_pose):
138
- starting_lane = edge
139
- break
140
-
141
- distance = edge.polygon.distance(Point(*ego_pose.array))
142
- if distance < closest_distance:
143
- starting_lane = edge
144
- closest_distance = distance
145
-
146
- return starting_lane
147
-
148
-
149
- def get_discrete_centerline(
150
- current_lane: LaneGraphEdgeMapObject,
151
- route_roadblock_dict: Dict[str, RoadBlockGraphEdgeMapObject],
152
- route_lane_dict: Dict[str, LaneGraphEdgeMapObject],
153
- search_depth: int = 30,
154
- ) -> List[StateSE2]:
155
- """
156
- Given the current lane, apply graph search, and extract centerline.
157
- :param current_lane: Lane object closest to ego
158
- :param route_roadblock_dict: Dictionary of roadblock ids and objects
159
- :param route_lane_dict: Dictionary of lane ids and objects
160
- :param search_depth: max search depth of Dijkstra, defaults to 30
161
- :return: List of (x, y, heading) objects
162
- """
163
-
164
- roadblocks = list(route_roadblock_dict.values())
165
- roadblock_ids = list(route_roadblock_dict.keys())
166
-
167
- # find current roadblock index
168
- start_idx = np.argmax(np.array(roadblock_ids) == current_lane.get_roadblock_id())
169
- roadblock_window = roadblocks[start_idx : start_idx + search_depth]
170
-
171
- graph_search = Dijkstra(current_lane, list(route_lane_dict.keys()))
172
- route_plan, _ = graph_search.search(roadblock_window[-1])
173
-
174
- centerline_discrete_path: List[StateSE2] = []
175
- for lane in route_plan:
176
- centerline_discrete_path.extend(lane.baseline_path.discrete_path)
177
-
178
- return centerline_discrete_path
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
openscene-v1.1/process_data/helpers/__init__.py DELETED
File without changes
openscene-v1.1/process_data/helpers/canbus.py DELETED
@@ -1,50 +0,0 @@
1
- import numpy as np
2
-
3
-
4
- class CanBus:
5
- """Wrapper class to convert lidar_can_bus to numpy array"""
6
-
7
- def __init__(self, lidar_pc):
8
- self.x = lidar_pc.ego_pose.x
9
- self.y = lidar_pc.ego_pose.y
10
- self.z = lidar_pc.ego_pose.z
11
-
12
- self.qw = lidar_pc.ego_pose.qw
13
- self.qx = lidar_pc.ego_pose.qx
14
- self.qy = lidar_pc.ego_pose.qy
15
- self.qz = lidar_pc.ego_pose.qz
16
-
17
- self.acceleration_x = lidar_pc.ego_pose.acceleration_x
18
- self.acceleration_y = lidar_pc.ego_pose.acceleration_y
19
- self.acceleration_z = lidar_pc.ego_pose.acceleration_z
20
-
21
- self.vx = lidar_pc.ego_pose.vx
22
- self.vy = lidar_pc.ego_pose.vy
23
- self.vz = lidar_pc.ego_pose.vz
24
-
25
- self.angular_rate_x = lidar_pc.ego_pose.angular_rate_x
26
- self.angular_rate_y = lidar_pc.ego_pose.angular_rate_y
27
- self.angular_rate_z = lidar_pc.ego_pose.angular_rate_z
28
-
29
- self.tensor = np.array(
30
- [
31
- self.x,
32
- self.y,
33
- self.z,
34
- self.qw,
35
- self.qx,
36
- self.qy,
37
- self.qz,
38
- self.acceleration_x,
39
- self.acceleration_y,
40
- self.acceleration_z,
41
- self.vx,
42
- self.vy,
43
- self.vz,
44
- self.angular_rate_x,
45
- self.angular_rate_y,
46
- self.angular_rate_z,
47
- 0.0,
48
- 0.0,
49
- ]
50
- )
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
openscene-v1.1/process_data/helpers/multiprocess_helper.py DELETED
@@ -1,32 +0,0 @@
1
- import numpy as np
2
- import multiprocessing
3
-
4
-
5
- def get_scenes_per_thread(scenes, thread_num):
6
- scenes = sorted(scenes)
7
- num_tasks = thread_num
8
- cur_id = int(multiprocessing.current_process().name)
9
-
10
- num_scene = len(scenes)
11
- a = num_scene // num_tasks
12
- b = num_scene % num_tasks
13
-
14
- if cur_id == 0:
15
- print("num_scene:", num_scene)
16
-
17
- process_num = []
18
- for id in range(num_tasks):
19
- if id >= b:
20
- process_num.append(a)
21
- else:
22
- process_num.append(a + 1)
23
- addsum = np.cumsum(process_num)
24
-
25
- if cur_id == 0:
26
- start = 0
27
- end = addsum[0]
28
- else:
29
- start = addsum[cur_id - 1]
30
- end = addsum[cur_id]
31
-
32
- return scenes[start:end], start
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
openscene-v1.1/process_data/helpers/multisweep_helper.py DELETED
@@ -1,66 +0,0 @@
1
- import os
2
- import numpy as np
3
- from pyquaternion import Quaternion
4
-
5
- from nuplan.database.nuplan_db_orm.lidar_pc import LidarPc
6
- from nuplan.database.nuplan_db_orm.lidar import Lidar
7
-
8
- from navsim.common.extraction.helpers.canbus import CanBus
9
-
10
-
11
- def obtain_sensor2top(lidar_token, log_db, l2e_t, l2e_r_mat, e2g_t, e2g_r_mat, args):
12
- """Obtain the info with RT matric from other sensors to Top LiDAR.
13
-
14
- Args:
15
- lidar_token (str): Sample data token corresponding to the
16
- specific sensor type.
17
- log_db: To obtain LiDAR of corresponding token.
18
- l2e_t (np.ndarray): Translation from lidar to ego in shape (1, 3).
19
- l2e_r_mat (np.ndarray): Rotation matrix from lidar to ego
20
- in shape (3, 3).
21
- e2g_t (np.ndarray): Translation from ego to global in shape (1, 3).
22
- e2g_r_mat (np.ndarray): Rotation matrix from ego to global
23
- in shape (3, 3).
24
-
25
- Returns:
26
- sweep (dict): Sweep information after transformation.
27
- """
28
- lidar_pc = log_db.session.query(LidarPc).filter(LidarPc.token == lidar_token).all()
29
- lidar_pc = lidar_pc[0]
30
- can_bus = CanBus(lidar_pc).tensor
31
-
32
- lidar_sensor = log_db.session.query(Lidar).filter(Lidar.token == lidar_pc.lidar_token).all()
33
- lidar_sensor = lidar_sensor[0]
34
-
35
- sweep = {
36
- "prev_sweep_token": lidar_pc.prev_token,
37
- "data_path": os.path.join(args.nuplan_sensor_path, lidar_pc.filename),
38
- "type": lidar_sensor.channel,
39
- "sample_data_token": lidar_pc.token,
40
- "sensor2ego_translation": lidar_sensor.translation_np,
41
- "sensor2ego_rotation": lidar_sensor.quaternion,
42
- "ego2global_translation": can_bus[:3],
43
- "ego2global_rotation": can_bus[3:7],
44
- "timestamp": lidar_pc.timestamp,
45
- }
46
-
47
- l2e_r_s = sweep["sensor2ego_rotation"]
48
- l2e_t_s = sweep["sensor2ego_translation"]
49
- e2g_r_s = sweep["ego2global_rotation"]
50
- e2g_t_s = sweep["ego2global_translation"]
51
-
52
- # obtain the RT from sensor to Top LiDAR
53
- # sweep->ego->global->ego'->lidar
54
- l2e_r_s_mat = Quaternion(l2e_r_s).rotation_matrix
55
- e2g_r_s_mat = Quaternion(e2g_r_s).rotation_matrix
56
- R = (l2e_r_s_mat.T @ e2g_r_s_mat.T) @ (np.linalg.inv(e2g_r_mat).T @ np.linalg.inv(l2e_r_mat).T)
57
- T = (l2e_t_s @ e2g_r_s_mat.T + e2g_t_s) @ (
58
- np.linalg.inv(e2g_r_mat).T @ np.linalg.inv(l2e_r_mat).T
59
- )
60
- T -= (
61
- e2g_t @ (np.linalg.inv(e2g_r_mat).T @ np.linalg.inv(l2e_r_mat).T)
62
- + l2e_t @ np.linalg.inv(l2e_r_mat).T
63
- )
64
- sweep["sensor2lidar_rotation"] = R.T # points @ R.T + T
65
- sweep["sensor2lidar_translation"] = T
66
- return sweep
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
openscene-v1.1/process_data/helpers/transformation.py DELETED
@@ -1,138 +0,0 @@
1
- import numpy as np
2
- from pyquaternion import Quaternion
3
-
4
- from nuplan.database.utils.pointclouds.pointcloud import PointCloud
5
-
6
-
7
- def _load_points(pc_file_name):
8
- pc = PointCloud.parse_from_file(pc_file_name).to_pcd_bin2().T
9
- return pc
10
-
11
-
12
- def transform_pcs_to_images(
13
- pc,
14
- cam2lidar_rotation,
15
- cam2lidar_translation,
16
- cam_intrinsic,
17
- img_shape=None,
18
- eps=1e-3,
19
- return_depth=False,
20
- ):
21
- """Transform point clouds from LiDAR coordinates to the camera coordinates.
22
-
23
- Args:
24
- pc: a numpy array with shape [-1, 6]
25
- cam_infos: dict of camera information.
26
- Return:
27
- pc_cam: dict of 2d coordinates in corresponding camera space.
28
- """
29
- pc_xyz = pc[:, :3]
30
-
31
- lidar2cam_r = np.linalg.inv(cam2lidar_rotation)
32
- lidar2cam_t = cam2lidar_translation @ lidar2cam_r.T
33
- lidar2cam_rt = np.eye(4)
34
- lidar2cam_rt[:3, :3] = lidar2cam_r.T
35
- lidar2cam_rt[3, :3] = -lidar2cam_t
36
-
37
- viewpad = np.eye(4)
38
- viewpad[: cam_intrinsic.shape[0], : cam_intrinsic.shape[1]] = cam_intrinsic
39
- lidar2img_rt = viewpad @ lidar2cam_rt.T
40
-
41
- cur_pc_xyz = np.concatenate([pc_xyz, np.ones_like(pc_xyz)[:, :1]], -1)
42
- cur_pc_cam = lidar2img_rt @ cur_pc_xyz.T
43
- cur_pc_cam = cur_pc_cam.T
44
-
45
- cur_pc_in_fov = cur_pc_cam[:, 2] > eps
46
- depth = cur_pc_cam[..., 2:3]
47
-
48
- cur_pc_cam = cur_pc_cam[..., 0:2] / np.maximum(
49
- cur_pc_cam[..., 2:3], np.ones_like(cur_pc_cam[..., 2:3]) * eps
50
- )
51
- if img_shape is not None:
52
- img_h, img_w = img_shape
53
- cur_pc_in_fov = (
54
- cur_pc_in_fov
55
- & (cur_pc_cam[:, 0] < (img_w - 1))
56
- & (cur_pc_cam[:, 0] > 0)
57
- & (cur_pc_cam[:, 1] < (img_h - 1))
58
- & (cur_pc_cam[:, 1] > 0)
59
- )
60
- if return_depth:
61
- cur_pc_cam = np.concatenate([cur_pc_cam, depth], axis=-1)
62
- return cur_pc_cam, cur_pc_in_fov
63
-
64
-
65
- def transform_cam_to_img(pc_cam, cam_intrinsic, img_shape=None, eps=1e-3, return_depth=False):
66
- """Transform point clouds from LiDAR coordinates to the camera coordinates.
67
-
68
- Args:
69
- pc: a numpy array with shape [-1, 6]
70
- cam_infos: dict of camera information.
71
- Return:
72
- pc_cam: dict of 2d coordinates in corresponding camera space.
73
- """
74
- pc_cam = pc_cam[:, :3]
75
-
76
- viewpad = np.eye(4)
77
- viewpad[: cam_intrinsic.shape[0], : cam_intrinsic.shape[1]] = cam_intrinsic
78
-
79
- pc_img = np.concatenate([pc_cam, np.ones_like(pc_cam)[:, :1]], -1)
80
- pc_img = viewpad @ pc_img.T
81
- pc_img = pc_img.T
82
-
83
- cur_pc_in_fov = pc_img[:, 2] > eps
84
- depth = pc_img[..., 2:3]
85
-
86
- pc_img = pc_img[..., 0:2] / np.maximum(pc_img[..., 2:3], np.ones_like(pc_img[..., 2:3]) * eps)
87
- if img_shape is not None:
88
- img_h, img_w = img_shape
89
- cur_pc_in_fov = (
90
- cur_pc_in_fov
91
- & (pc_img[:, 0] < (img_w - 1))
92
- & (pc_img[:, 0] > 0)
93
- & (pc_img[:, 1] < (img_h - 1))
94
- & (pc_img[:, 1] > 0)
95
- )
96
- if return_depth:
97
- pc_img = np.concatenate([pc_img, depth], axis=-1)
98
- return pc_img, cur_pc_in_fov
99
-
100
-
101
- def transform_nuplan_boxes_to_cam(box, cam2lidar_rotation, cam2lidar_translation):
102
- """Transform point clouds from LiDAR coordinates to the camera coordinates.
103
-
104
- Args:
105
- box: a numpy array with shape [-1, 7]
106
- """
107
- locs, dims, rots = box[:, :3], box[:, 3:6], box[:, 6:]
108
- dims_cams = dims[:, [0, 2, 1]] # l, w, h -> l, h, w
109
-
110
- rots_cam = np.zeros_like(rots)
111
- for idx, rot in enumerate(rots):
112
- rot = Quaternion(axis=[0, 0, 1], radians=rot)
113
- rot = Quaternion(matrix=cam2lidar_rotation).inverse * rot
114
- rots_cam[idx] = -rot.yaw_pitch_roll[0]
115
-
116
- lidar2cam_r = np.linalg.inv(cam2lidar_rotation)
117
- lidar2cam_t = cam2lidar_translation @ lidar2cam_r.T
118
- lidar2cam_rt = np.eye(4)
119
- lidar2cam_rt[:3, :3] = lidar2cam_r.T
120
- lidar2cam_rt[3, :3] = -lidar2cam_t
121
-
122
- locs_cam = np.concatenate([locs, np.ones_like(locs)[:, :1]], -1) # -1, 4
123
- locs_cam = lidar2cam_rt.T @ locs_cam.T
124
- locs_cam = locs_cam.T
125
- locs_cam = locs_cam[:, :-1]
126
- return np.concatenate([locs_cam, dims_cams, rots_cam], -1)
127
-
128
-
129
- def transform_sweep_pc_to_lidar_top(sweep_pc, sensor2lidar_rotation, sensor2lidar_translation):
130
- sweep_xyz = sweep_pc[:, :3]
131
- sweep_xyz = np.concatenate([sweep_xyz, np.ones_like(sweep_xyz)[:, :1]], -1)
132
-
133
- sensor2lidar_rt = np.eye(4)
134
- sensor2lidar_rt[:3, :3] = sensor2lidar_rotation.T
135
- sensor2lidar_rt[3, :3] = sensor2lidar_translation
136
-
137
- sweep_xyz = sweep_xyz @ sensor2lidar_rt
138
- return sweep_xyz[:, :3]
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
openscene-v1.1/process_data/helpers/viz.py DELETED
@@ -1,93 +0,0 @@
1
- import cv2
2
- import numpy as np
3
-
4
- from navsim.common.extraction.helpers import transformation
5
-
6
-
7
- def draw_pcs_on_images(pc, cam_infos, eps=1e-3):
8
- for cam_type, cam_info in cam_infos.items():
9
- cur_img_path = cam_info["data_path"]
10
- cur_img = cv2.imread(cur_img_path)
11
- cur_img_h, cur_img_w = cur_img.shape[:2]
12
-
13
- cur_pc_cam, cur_pc_in_fov = transformation.transform_pcs_to_images(
14
- pc,
15
- cam_info["sensor2lidar_rotation"],
16
- cam_info["sensor2lidar_translation"],
17
- cam_info["cam_intrinsic"],
18
- img_shape=(cur_img_h, cur_img_w),
19
- eps=eps,
20
- )
21
-
22
- cur_pc_cam = cur_pc_cam[cur_pc_in_fov]
23
- for x, y in cur_pc_cam:
24
- cv2.circle(cur_img, (int(x), int(y)), 3, (255, 0, 0), 3)
25
- cv2.imwrite(f"dbg/{cam_type}.png", cur_img)
26
- return None
27
-
28
-
29
- def draw_pcs(points, labels, color_map):
30
- """Draw point cloud from BEV
31
-
32
- Args:
33
- points: A ndarray with shape as [-1, 3]
34
- labels: the label of each point with shape [-1]
35
- color_map: color of each label.
36
- """
37
- import matplotlib.pyplot as plt
38
-
39
- _, ax = plt.subplots(1, 1, figsize=(9, 9))
40
- axes_limit = 40
41
- # points: LiDAR points with shape [-1, 3]
42
- viz_points = points
43
- dists = np.sqrt(np.sum(viz_points[:, :2] ** 2, axis=1))
44
- colors = np.minimum(1, dists / axes_limit / np.sqrt(2))
45
-
46
- # prepare color_map
47
- points_color = color_map[labels] / 255.0 # -1, 3
48
-
49
- point_scale = 0.2
50
- scatter = ax.scatter(viz_points[:, 0], viz_points[:, 1], c=points_color, s=point_scale)
51
-
52
- ax.plot(0, 0, "x", color="red")
53
- ax.set_xlim(-axes_limit, axes_limit)
54
- ax.set_ylim(-axes_limit, axes_limit)
55
- ax.axis("off")
56
- ax.set_aspect("equal")
57
- plt.savefig("dbg/dbg.png", bbox_inches="tight", pad_inches=0, dpi=200)
58
-
59
-
60
- def draw_sweep_pcs(info, sweeps):
61
- cur_pc_file = info["lidar_path"]
62
- cur_pc = transformation._load_points(cur_pc_file)
63
- viz_pcs = [cur_pc[:, :3]]
64
- viz_labels = [np.ones_like(cur_pc)[:, 0] * 0.0]
65
-
66
- for idx, sweep in enumerate(sweeps):
67
- sweep_pc_file = sweep["data_path"]
68
- sweep_pc = transformation._load_points(sweep_pc_file)
69
- sweep_pc = transformation.transform_sweep_pc_to_lidar_top(
70
- sweep_pc, sweep["sensor2lidar_rotation"], sweep["sensor2lidar_translation"]
71
- )
72
-
73
- viz_pcs.append(sweep_pc)
74
- viz_labels.append(np.ones_like(sweep_pc)[:, 0] * (idx + 1))
75
-
76
- viz_pcs = np.concatenate(viz_pcs, 0)
77
- viz_labels = np.concatenate(viz_labels, 0).astype(np.int)
78
- color_map = np.array(
79
- [
80
- [245, 150, 100],
81
- [245, 230, 100],
82
- [250, 80, 100],
83
- [150, 60, 30],
84
- [255, 0, 0],
85
- [180, 30, 80],
86
- [255, 0, 0],
87
- [30, 30, 255],
88
- [200, 40, 255],
89
- [90, 30, 150],
90
- ]
91
- )
92
- draw_pcs(viz_pcs, viz_labels, color_map)
93
- return None
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
openscene-v1.1/process_data/helpers/viz_box.py DELETED
@@ -1,102 +0,0 @@
1
- import numpy as np
2
-
3
- from matplotlib.lines import Line2D
4
-
5
-
6
- def rotz(t):
7
- """Rotation about the z-axis."""
8
- c = np.cos(t)
9
- s = np.sin(t)
10
- return np.array([[c, -s, 0], [s, c, 0], [0, 0, 1]])
11
-
12
-
13
- def draw_box_and_pc(points, boxes, names=None):
14
- import matplotlib.pyplot as plt
15
-
16
- _, ax = plt.subplots(1, 1, figsize=(9, 9))
17
- axes_limit = 40
18
- # points: LiDAR points with shape [-1, 3]
19
- viz_points = points
20
- dists = np.sqrt(np.sum(viz_points[:, :2] ** 2, axis=1))
21
- colors = np.minimum(1, dists / axes_limit / np.sqrt(2))
22
-
23
- # prepare color_map
24
- points_color = np.array([[245, 150, 100]]).reshape(1, 3).repeat(points.shape[0], 0)
25
- points_color = points_color / 255.0
26
- point_scale = 0.2
27
- scatter = ax.scatter(viz_points[:, 0], viz_points[:, 1], c=points_color, s=point_scale)
28
-
29
- ax.plot(0, 0, "x", color="red")
30
- ax.set_xlim(-axes_limit, axes_limit)
31
- ax.set_ylim(-axes_limit, axes_limit)
32
-
33
- if names is None:
34
- draw_box_3d(boxes, ax)
35
- else:
36
- color_map = {
37
- "vehicle": "green",
38
- "bicycle": "blue",
39
- "pedestrian": "red",
40
- "traffic_cone": "darkred",
41
- "barrier": "peru",
42
- "czone_sign": "pink",
43
- "generic_object": "black",
44
- }
45
- colors = [color_map[name] for name in names]
46
- draw_box_3d(boxes, ax, linestyle="--", colors=colors)
47
-
48
- ax.axis("off")
49
- ax.set_aspect("equal")
50
- plt.savefig("dbg/dbg.png", bbox_inches="tight", pad_inches=0, dpi=200)
51
-
52
-
53
- def draw_box_3d(boxes, ax, linestyle="--", colors="cyan"):
54
- for o in range(len(boxes)):
55
- color = colors[o]
56
- obj = boxes[o]
57
- c_x, c_y, c_z, l, w, h, rz = obj[0], obj[1], obj[2], obj[3], obj[4], obj[5], obj[6]
58
-
59
- R = rotz(rz)
60
-
61
- # 3d bounding box corners
62
- x_corners = [l / 2, l / 2, -l / 2, -l / 2, l / 2, l / 2, -l / 2, -l / 2]
63
- y_corners = [w / 2, -w / 2, -w / 2, w / 2, w / 2, -w / 2, -w / 2, w / 2]
64
- z_corners = [h, h, h, h, 0, 0, 0, 0]
65
-
66
- # rotate and translate 3d bounding box
67
- corners_3d = np.dot(R, np.vstack([x_corners, y_corners, z_corners]))
68
-
69
- corners_3d[0, :] = corners_3d[0, :] + c_x
70
- corners_3d[1, :] = corners_3d[1, :] + c_y
71
- corners_3d[2, :] = corners_3d[2, :] + c_z
72
-
73
- corners_3d_velo = np.transpose(corners_3d)
74
-
75
- x1, x2, x3, x4 = corners_3d_velo[0:4, 0]
76
- y1, y2, y3, y4 = corners_3d_velo[0:4, 1]
77
-
78
- polygon = np.zeros([5, 2], dtype=np.float32)
79
- polygon[0, 0] = x1
80
- polygon[1, 0] = x2
81
- polygon[2, 0] = x3
82
- polygon[3, 0] = x4
83
- polygon[4, 0] = x1
84
-
85
- polygon[0, 1] = y1
86
- polygon[1, 1] = y2
87
- polygon[2, 1] = y3
88
- polygon[3, 1] = y4
89
- polygon[4, 1] = y1
90
-
91
- line1 = [(x1, y1), (x2, y2)]
92
- line2 = [(x2, y2), (x3, y3)]
93
- line3 = [(x3, y3), (x4, y4)]
94
- line4 = [(x4, y4), (x1, y1)]
95
- (line1_xs, line1_ys) = zip(*line1)
96
- (line2_xs, line2_ys) = zip(*line2)
97
- (line3_xs, line3_ys) = zip(*line3)
98
- (line4_xs, line4_ys) = zip(*line4)
99
- ax.add_line(Line2D(line1_xs, line1_ys, linewidth=1, linestyle=linestyle, color=color))
100
- ax.add_line(Line2D(line2_xs, line2_ys, linewidth=1, linestyle=linestyle, color=color))
101
- ax.add_line(Line2D(line3_xs, line3_ys, linewidth=1, linestyle=linestyle, color=color))
102
- ax.add_line(Line2D(line4_xs, line4_ys, linewidth=1, linestyle=linestyle, color=color))
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
openscene-v1.1/process_data/helpers/viz_box_2d.py DELETED
@@ -1,135 +0,0 @@
1
- """
2
- Visualize 3D boxes in Image space.
3
- Align the setting in mmdetection3d:
4
- * Convert 3D box in nuplan coordinates to camera coordinates.
5
- * draw 3D box in camera.
6
- """
7
-
8
- import cv2
9
- import numpy as np
10
- from navsim.common.extraction.helpers import transformation
11
-
12
-
13
- def rotation_3d_in_axis(points, angles, axis=0):
14
- """Rotate points by angles according to axis.
15
-
16
- Args:
17
- points (torch.Tensor): Points of shape (N, M, 3).
18
- angles (torch.Tensor): Vector of angles in shape (N,)
19
- axis (int, optional): The axis to be rotated. Defaults to 0.
20
-
21
- Raises:
22
- ValueError: when the axis is not in range [0, 1, 2], it will \
23
- raise value error.
24
-
25
- Returns:
26
- torch.Tensor: Rotated points in shape (N, M, 3)
27
- """
28
- rot_sin = np.sin(angles)
29
- rot_cos = np.cos(angles)
30
- ones = np.ones_like(rot_cos)
31
- zeros = np.zeros_like(rot_cos)
32
- if axis == 1:
33
- rot_mat_T = np.stack(
34
- [
35
- np.stack([rot_cos, zeros, -rot_sin]),
36
- np.stack([zeros, ones, zeros]),
37
- np.stack([rot_sin, zeros, rot_cos]),
38
- ]
39
- )
40
- elif axis == 2 or axis == -1:
41
- rot_mat_T = np.stack(
42
- [
43
- np.stack([rot_cos, -rot_sin, zeros]),
44
- np.stack([rot_sin, rot_cos, zeros]),
45
- np.stack([zeros, zeros, ones]),
46
- ]
47
- )
48
- elif axis == 0:
49
- rot_mat_T = np.stack(
50
- [
51
- np.stack([zeros, rot_cos, -rot_sin]),
52
- np.stack([zeros, rot_sin, rot_cos]),
53
- np.stack([ones, zeros, zeros]),
54
- ]
55
- )
56
- else:
57
- raise ValueError(f"axis should in range [0, 1, 2], got {axis}")
58
- return np.einsum("aij,jka->aik", points, rot_mat_T)
59
-
60
-
61
- def plot_rect3d_on_img(img, num_rects, rect_corners, color=(0, 255, 0), thickness=1):
62
- """Plot the boundary lines of 3D rectangular on 2D images.
63
-
64
- Args:
65
- img (numpy.array): The numpy array of image.
66
- num_rects (int): Number of 3D rectangulars.
67
- rect_corners (numpy.array): Coordinates of the corners of 3D
68
- rectangulars. Should be in the shape of [num_rect, 8, 2].
69
- color (tuple[int]): The color to draw bboxes. Default: (0, 255, 0).
70
- thickness (int, optional): The thickness of bboxes. Default: 1.
71
- """
72
- line_indices = (
73
- (0, 1),
74
- (0, 3),
75
- (0, 4),
76
- (1, 2),
77
- (1, 5),
78
- (3, 2),
79
- (3, 7),
80
- (4, 5),
81
- (4, 7),
82
- (2, 6),
83
- (5, 6),
84
- (6, 7),
85
- )
86
- for i in range(num_rects):
87
- corners = rect_corners[i].astype(np.int)
88
- for start, end in line_indices:
89
- cv2.line(
90
- img,
91
- (corners[start, 0], corners[start, 1]),
92
- (corners[end, 0], corners[end, 1]),
93
- color,
94
- thickness,
95
- cv2.LINE_AA,
96
- )
97
- return img.astype(np.uint8)
98
-
99
-
100
- def draw_boxes_nuplan_on_img(gt_boxes_nuplan, cam_infos, eps=1e-3):
101
- for cam_type, cam_info in cam_infos.items():
102
- cur_img_path = cam_info["data_path"]
103
- cur_img = cv2.imread(cur_img_path)
104
- cur_img_h, cur_img_w = cur_img.shape[:2]
105
-
106
- gt_boxes_cams = transformation.transform_nuplan_boxes_to_cam(
107
- gt_boxes_nuplan,
108
- cam_info["sensor2lidar_rotation"],
109
- cam_info["sensor2lidar_translation"],
110
- )
111
-
112
- # Then convert gt_boxes_cams to corners.
113
- cur_locs, cur_dims, cur_rots = (
114
- gt_boxes_cams[:, :3],
115
- gt_boxes_cams[:, 3:6],
116
- gt_boxes_cams[:, 6:],
117
- )
118
- corners_norm = np.stack(np.unravel_index(np.arange(8), [2] * 3), axis=1)
119
- corners_norm = corners_norm[[0, 1, 3, 2, 4, 5, 7, 6]]
120
- corners_norm = corners_norm - np.array([0.5, 0.5, 0.5])
121
- corners = cur_dims.reshape([-1, 1, 3]) * corners_norm.reshape([1, 8, 3])
122
- corners = rotation_3d_in_axis(corners, cur_rots.squeeze(-1), axis=1)
123
- corners += cur_locs.reshape(-1, 1, 3)
124
-
125
- # Then draw project corners to image.
126
- corners_img, corners_pc_in_fov = transformation.transform_cam_to_img(
127
- corners.reshape(-1, 3), cam_info["cam_intrinsic"], img_shape=(cur_img_h, cur_img_w)
128
- )
129
- corners_img = corners_img.reshape(-1, 8, 2)
130
- corners_pc_in_fov = corners_pc_in_fov.reshape(-1, 8)
131
- valid_corners = corners_pc_in_fov.all(-1)
132
- corners_img = corners_img[valid_corners]
133
- cur_img = plot_rect3d_on_img(cur_img, len(corners_img), corners_img)
134
- cv2.imwrite(f"dbg/{cam_type}.png", cur_img)
135
- return None
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
openscene-v1.1/process_data/process.sh DELETED
@@ -1,32 +0,0 @@
1
- # Scripts for processing nuplan data.
2
- # export PYTHONPATH=/cpfs01/user/yangzetong/code/workshop_codes/nuplan-devkit:${PYTHONPATH}
3
-
4
-
5
- # export NUPLAN_DATA_ROOT="$HOME/nuplan/dataset/"
6
- # export NUPLAN_MAPS_ROOT="$HOME/nuplan/dataset/maps"
7
- # export NUPLAN_EXP_ROOT="$HOME/nuplan/exp"
8
- # export NUPLAN_DEVKIT_ROOT="$HOME/nuplan-devkit/"
9
-
10
- # 0. Design nuPlan data path, and output data path.
11
- NUPLAN_PATH=$HOME/nuplan/dataset/nuplan-v1.1/
12
- NUPLAN_DB_PATH=$NUPLAN_PATH/splits/mini
13
- NUPLAN_SENSOR_PATH=${NUPLAN_PATH}/sensor_blobs
14
- NUPLAN_MAP_VERSION=nuplan-maps-v1.0
15
- NUPLAN_MAPS_ROOT=${HOME}/nuplan/dataset/maps
16
-
17
- OUT_DIR=/home/daniel/navsim_logs/mini
18
-
19
- # 1. TODO: Generate train/val pickle.
20
- python $NAVSIM_DEVKIT_ROOT/navsim/common/extraction/create_nuplan_data_with_vis.py \
21
- --nuplan-root-path ${NUPLAN_PATH} \
22
- --nuplan-db-path ${NUPLAN_DB_PATH} \
23
- --nuplan-sensor-path ${NUPLAN_SENSOR_PATH} \
24
- --nuplan-map-version ${NUPLAN_MAP_VERSION} \
25
- --nuplan-map-root ${NUPLAN_MAPS_ROOT} \
26
- --out-dir ${OUT_DIR} \
27
- --split mini \
28
- --thread-num 16
29
-
30
- # 2. TODO: Extract Data Files from the generated pickle file.
31
-
32
- # 3. TODO: Soft-link to some place similar to https://github.com/OpenDriveLab/OpenScene/blob/main/docs/dataset_stats.md#filesystem-hierarchy