-
Notifications
You must be signed in to change notification settings - Fork 47
/
Copy pathprocessors.py
199 lines (165 loc) · 8.83 KB
/
processors.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
from typing import List
import numpy as np
import tensorflow as tf
from tensorflow.python.keras.utils.data_utils import Sequence
from config import Parameters
from point_pillars import createPillars, createPillarsTarget
from readers import DataReader, Label3D
from sklearn.utils import shuffle
import sys
def select_best_anchors(arr):
dims = np.indices(arr.shape[1:])
# arr[..., 0:1] gets the occupancy value from occ in {-1, 0, 1}, i.e. {bad match, neg box, pos box}
ind = (np.argmax(arr[..., 0:1], axis=0),) + tuple(dims)
return arr[ind]
class DataProcessor(Parameters):
def __init__(self):
super(DataProcessor, self).__init__()
anchor_dims = np.array(self.anchor_dims, dtype=np.float32)
self.anchor_dims = anchor_dims[:, 0:3]
self.anchor_z = anchor_dims[:, 3]
self.anchor_yaw = anchor_dims[:, 4]
# Counts may be used to make statistic about how well the anchor boxes fit the objects
self.pos_cnt, self.neg_cnt = 0, 0
@staticmethod
def transform_labels_into_lidar_coordinates(labels: List[Label3D], R: np.ndarray, t: np.ndarray):
transformed = []
for label in labels:
label.centroid = label.centroid @ np.linalg.inv(R).T - t
label.dimension = label.dimension[[2, 1, 0]]
label.yaw -= np.pi / 2
while label.yaw < -np.pi:
label.yaw += (np.pi * 2)
while label.yaw > np.pi:
label.yaw -= (np.pi * 2)
transformed.append(label)
return labels
def make_point_pillars(self, points: np.ndarray):
assert points.ndim == 2
assert points.shape[1] == 4
assert points.dtype == np.float32
pillars, indices = createPillars(points,
self.max_points_per_pillar,
self.max_pillars,
self.x_step,
self.y_step,
self.x_min,
self.x_max,
self.y_min,
self.y_max,
self.z_min,
self.z_max,
False)
return pillars, indices
def make_ground_truth(self, labels: List[Label3D]):
# filter labels by classes (cars, pedestrians and Trams)
# Label has 4 properties (Classification (0th index of labels file),
# centroid coordinates, dimensions, yaw)
labels = list(filter(lambda x: x.classification in self.classes, labels))
if len(labels) == 0:
pX, pY = int(self.Xn / self.downscaling_factor), int(self.Yn / self.downscaling_factor)
a = int(self.anchor_dims.shape[0])
return np.zeros((pX, pY, a), dtype='float32'), np.zeros((pX, pY, a, self.nb_dims), dtype='float32'), \
np.zeros((pX, pY, a, self.nb_dims), dtype='float32'), np.zeros((pX, pY, a), dtype='float32'), \
np.zeros((pX, pY, a, self.nb_classes), dtype='float64')
# For each label file, generate these properties except for the Don't care class
target_positions = np.array([label.centroid for label in labels], dtype=np.float32)
target_dimension = np.array([label.dimension for label in labels], dtype=np.float32)
target_yaw = np.array([label.yaw for label in labels], dtype=np.float32)
target_class = np.array([self.classes[label.classification] for label in labels], dtype=np.int32)
assert np.all(target_yaw >= -np.pi) & np.all(target_yaw <= np.pi)
assert len(target_positions) == len(target_dimension) == len(target_yaw) == len(target_class)
target, pos, neg = createPillarsTarget(target_positions,
target_dimension,
target_yaw,
target_class,
self.anchor_dims,
self.anchor_z,
self.anchor_yaw,
self.positive_iou_threshold,
self.negative_iou_threshold,
self.nb_classes,
self.downscaling_factor,
self.x_step,
self.y_step,
self.x_min,
self.x_max,
self.y_min,
self.y_max,
self.z_min,
self.z_max,
False)
self.pos_cnt += pos
self.neg_cnt += neg
# return a merged target view for all objects in the ground truth and get categorical labels
sel = select_best_anchors(target)
ohe = tf.keras.utils.to_categorical(sel[..., 9], num_classes=self.nb_classes, dtype='float64')
return sel[..., 0], sel[..., 1:4], sel[..., 4:7], sel[..., 7], sel[..., 8], ohe
class SimpleDataGenerator(DataProcessor, Sequence):
""" Multiprocessing-safe data generator for training, validation or testing, without fancy augmentation """
def __init__(self, data_reader: DataReader, batch_size: int, lidar_files: List[str], label_files: List[str] = None,
calibration_files: List[str] = None):
super(SimpleDataGenerator, self).__init__()
self.data_reader = data_reader
self.batch_size = batch_size
self.lidar_files = lidar_files
self.label_files = label_files
self.calibration_files = calibration_files
assert (calibration_files is None and label_files is None) or \
(calibration_files is not None and label_files is not None)
if self.calibration_files is not None:
assert len(self.calibration_files) == len(self.lidar_files)
assert len(self.label_files) == len(self.lidar_files)
def __len__(self):
return len(self.lidar_files) // self.batch_size
def __getitem__(self, batch_id: int):
file_ids = np.arange(batch_id * self.batch_size, self.batch_size * (batch_id + 1))
# print("inside getitem")
pillars = []
voxels = []
occupancy = []
position = []
size = []
angle = []
heading = []
classification = []
for i in file_ids:
lidar = self.data_reader.read_lidar(self.lidar_files[i])
# For each file, dividing the space into a x-y grid to create pillars
# Voxels are the pillar ids
pillars_, voxels_ = self.make_point_pillars(lidar)
pillars.append(pillars_)
voxels.append(voxels_)
if self.label_files is not None:
label = self.data_reader.read_label(self.label_files[i])
R, t = self.data_reader.read_calibration(self.calibration_files[i])
# Labels are transformed into the lidar coordinate bounding boxes
# Label has 7 values, centroid, dimensions and yaw value.
label_transformed = self.transform_labels_into_lidar_coordinates(label, R, t)
# These definitions can be found in point_pillars.cpp file
# We are splitting a 10 dim vector that contains this information.
occupancy_, position_, size_, angle_, heading_, classification_ = self.make_ground_truth(
label_transformed)
occupancy.append(occupancy_)
position.append(position_)
size.append(size_)
angle.append(angle_)
heading.append(heading_)
classification.append(classification_)
pillars = np.concatenate(pillars, axis=0)
voxels = np.concatenate(voxels, axis=0)
if self.label_files is not None:
occupancy = np.array(occupancy)
position = np.array(position)
size = np.array(size)
angle = np.array(angle)
heading = np.array(heading)
classification = np.array(classification)
return [pillars, voxels], [occupancy, position, size, angle, heading, classification]
else:
return [pillars, voxels]
def on_epoch_end(self):
# print("inside epoch")
if self.label_files is not None:
self.lidar_files, self.label_files, self.calibration_files = \
shuffle(self.lidar_files, self.label_files, self.calibration_files)