-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathkitti_3d_detection.py
285 lines (257 loc) · 10.2 KB
/
kitti_3d_detection.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
from argparse import ArgumentParser
from configparser import ConfigParser
from pathlib import Path
import numpy as np
import torch
import tqdm
from torch.utils.data import DataLoader
from datasets.kitti_dataset import (KittiDataset, collate_fn,
write_kitti_detection_results)
from datasets.voxelization import VoxelGeneratorWrapper
from detection.voxel_rcnn.utils import rotate_points_bev
from detection.voxel_rcnn.voxel_rcnn import VoxelRCNN
def model_inference(
root_path,
output_dir,
model,
weight_path,
voxel_size,
point_cloud_range,
num_point_features,
max_num_points_per_voxel,
max_num_voxels,
class_names,
batch_size,
test_time_aug_cfgs=None,
):
weights = torch.load(weight_path)
model.load_state_dict(weights)
model.eval()
model.cuda()
voxel_generator = VoxelGeneratorWrapper(
voxel_size=voxel_size,
point_cloud_range=point_cloud_range,
num_point_features=num_point_features,
max_num_points_per_voxel=max_num_points_per_voxel,
max_num_voxels=max_num_voxels,
)
for seq in range(21):
seq = str(seq).zfill(4)
if test_time_aug_cfgs is None:
dataset = KittiDataset(
root_path=root_path,
seq=seq,
voxel_generator=voxel_generator,
point_cloud_range=point_cloud_range,
aug_cfg=None,
)
data_loader = DataLoader(
dataset, batch_size=batch_size, collate_fn=collate_fn
)
pbar = tqdm.tqdm(data_loader)
for batch_dict in pbar:
for key, value in batch_dict.items():
if key not in ["frame_id", "batch_size"]:
batch_dict[key] = torch.from_numpy(value).float().cuda()
with torch.no_grad():
batch_dict = model.forward(batch_dict)
pred_dicts = model.post_processing(batch_dict)
write_kitti_detection_results(
pred_dicts,
class_names=class_names,
frame_ids=batch_dict["frame_id"],
calibrations=dataset.calib,
img_hw_dict=dataset.img_hw_dict,
output_dir=output_dir / seq,
)
else:
# no tta
seq_middle_dir = output_dir / "middle_data" / "no_aug" / seq
seq_middle_dir.mkdir(exist_ok=True, parents=True)
dataset = KittiDataset(
root_path=root_path,
seq=seq,
voxel_generator=voxel_generator,
point_cloud_range=point_cloud_range,
aug_cfg=None,
)
data_loader = DataLoader(dataset, collate_fn=collate_fn)
pbar = tqdm.tqdm(data_loader)
for batch_dict in pbar:
pbar.set_description(f"{seq}-no_aug")
for key, value in batch_dict.items():
if key not in ["frame_id", "batch_size"]:
batch_dict[key] = torch.from_numpy(value).float().cuda()
with torch.no_grad():
batch_dict = model.forward(batch_dict)
pred_dict = {
"batch_size": batch_dict["batch_size"],
"batch_box_preds": batch_dict["batch_box_preds"],
"batch_cls_preds": batch_dict["batch_cls_preds"],
"frame_ids": batch_dict["frame_id"],
"calib": dataset.calib,
"class_names": class_names,
"aug_cfg": None,
}
if len(class_names) > 1:
pred_dict["roi_labels"] = batch_dict["roi_labels"]
torch.save(
pred_dict, seq_middle_dir / f"{'_'.join(batch_dict['frame_id'])}.pt"
)
# tta
for aug_name, aug_params in test_time_aug_cfgs.items():
for aug_idx, aug_param in enumerate(aug_params):
aug_cfg = (aug_name, aug_param)
seq_middle_dir = (
output_dir / "middle_data" / aug_name / str(aug_idx) / seq
)
seq_middle_dir.mkdir(exist_ok=True, parents=True)
dataset = KittiDataset(
root_path=root_path,
seq=seq,
voxel_generator=voxel_generator,
point_cloud_range=point_cloud_range,
aug_cfg=aug_cfg,
)
data_loader = DataLoader(dataset, collate_fn=collate_fn)
pbar = tqdm.tqdm(data_loader)
for batch_dict in pbar:
pbar.set_description(f"{seq}-{aug_name}-{aug_idx}")
for key, value in batch_dict.items():
if key not in ["frame_id", "batch_size"]:
batch_dict[key] = torch.from_numpy(value).float().cuda()
with torch.no_grad():
batch_dict = model.forward(batch_dict)
pred_dict = {
"batch_size": batch_dict["batch_size"],
"batch_box_preds": batch_dict["batch_box_preds"],
"batch_cls_preds": batch_dict["batch_cls_preds"],
"frame_ids": batch_dict["frame_id"],
"calib": dataset.calib,
"class_names": class_names,
"aug_cfg": aug_cfg,
}
if len(class_names) > 1:
pred_dict["roi_labels"] = batch_dict["roi_labels"]
torch.save(
pred_dict,
seq_middle_dir / f"{'_'.join(batch_dict['frame_id'])}.pt",
)
def reverse_aug_boxes(data_dict):
boxes = data_dict["batch_box_preds"]
aug_method, aug_param = data_dict["aug_cfg"]
if aug_method == "scale":
boxes[:, :, :6] /= aug_param
elif aug_method == "rotate":
boxes = rotate_points_bev(
boxes, boxes.new_tensor(-aug_param).repeat(len(boxes), 1)
)
boxes[:, :, 6] -= aug_param
elif aug_method == "translate":
boxes[:, :, :3] -= boxes.new_tensor(aug_param)
elif aug_method == "flip":
boxes[:, :, 1] *= -1
boxes[:, :, 6] *= -1
data_dict["batch_box_preds"] = boxes
return data_dict
def tta_fusion(
root_path, output_dir: Path, model, class_names, test_time_aug_cfgs: dict
):
for seq in range(21):
seq = str(seq).zfill(4)
middle_dir = output_dir / "middle_data"
no_aug_dir = middle_dir / "no_aug" / seq
dataset = KittiDataset(
root_path=root_path,
seq=seq,
voxel_generator=None,
point_cloud_range=None,
aug_cfg=None,
)
pbar = tqdm.tqdm(list(no_aug_dir.iterdir()))
pbar.set_description(f"fusing {seq} data")
for filepath in pbar:
filename = filepath.name
data_dicts = [torch.load(filepath)]
for aug_name, aug_params in test_time_aug_cfgs.items():
for aug_idx in range(len(aug_params)):
cur_aug_data = torch.load(
middle_dir / aug_name / str(aug_idx) / seq / filename
)
cur_aug_data = reverse_aug_boxes(cur_aug_data)
data_dicts.append(cur_aug_data)
batch_box_preds = torch.cat(
[d["batch_box_preds"] for d in data_dicts], dim=1
)
batch_cls_preds = torch.cat(
[d["batch_cls_preds"] for d in data_dicts], dim=1
)
data_dict = data_dicts[0]
data_dict["batch_box_preds"] = batch_box_preds
data_dict["batch_cls_preds"] = batch_cls_preds
pred_dicts = model.post_processing(data_dict)
write_kitti_detection_results(
pred_dicts,
class_names=class_names,
frame_ids=data_dicts[0]["frame_ids"],
calibrations=dataset.calib,
img_hw_dict=dataset.img_hw_dict,
output_dir=output_dir / seq,
)
def main():
parser = ArgumentParser()
parser.add_argument("config", type=str)
parser.add_argument("split", type=str)
parser.add_argument("--inference", action="store_true")
args = parser.parse_args()
config = ConfigParser()
config.read(args.config)
root_dir = Path(config["data"]["root_dir"]) / args.split
det3d_name = config["detection"]["det3d_name"]
output_dir = root_dir / "det3d_out" / det3d_name
weight_path = config["detection"]["det3d_ckpt"]
batch_size = config["detection"].getint("batch_size")
score_thresh = config["detection"].getfloat("score_thresh")
tta = config["detection"].getboolean("tta")
test_time_aug_cfgs = (
{
"rotate": [np.pi / 4, -np.pi / 4, np.pi / 8, -np.pi / 8],
"scale": [0.95, 1.05],
"translate": [[-5, 0, 0], [5, 0, 0], [0, -5, 0], [0, 5, 0]],
"flip": [None],
}
if tta
else None
)
point_cloud_range = [0, -40, -3, 70.4, 40, 1]
voxel_size = [0.05, 0.05, 0.1]
num_point_features = 4
max_num_points_per_voxel = 5
max_num_voxels = 40000
class_names = ["Car"]
model = VoxelRCNN(
input_channels=num_point_features,
num_classes=len(class_names),
point_cloud_range=point_cloud_range,
voxel_size=voxel_size,
score_thresh=score_thresh,
)
if args.inference:
model_inference(
root_dir,
output_dir,
model,
weight_path,
voxel_size,
point_cloud_range,
num_point_features,
max_num_points_per_voxel,
max_num_voxels,
class_names,
batch_size,
test_time_aug_cfgs,
)
if tta:
tta_fusion(root_dir, output_dir, model, class_names, test_time_aug_cfgs)
if __name__ == "__main__":
main()