-
Notifications
You must be signed in to change notification settings - Fork 47
/
inference_utils.py
134 lines (114 loc) · 6.03 KB
/
inference_utils.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
import numpy as np
import cv2 as cv
from typing import List
from config import Parameters
from readers import DataReader
from processors import DataProcessor
class BBox(tuple):
""" bounding box tuple that can easily be accessed while being compatible to cv2 rotational rects """
def __new__(cls, bb_x, bb_y, bb_z, bb_length, bb_width, bb_height, bb_yaw, bb_heading, bb_cls, bb_conf):
bbx_tuple = ((float(bb_x), float(bb_y)), (float(bb_length), float(bb_width)), float(np.rad2deg(bb_yaw)))
return super(BBox, cls).__new__(cls, tuple(bbx_tuple))
def __init__(self, bb_x, bb_y, bb_z, bb_length, bb_width, bb_height, bb_yaw, bb_heading, bb_cls, bb_conf):
self.x = bb_x
self.y = bb_y
self.z = bb_z
self.length = bb_length
self.width = bb_width
self.height = bb_height
self.yaw = bb_yaw
self.heading = bb_heading
self.cls = bb_cls
self.conf = bb_conf
def __str__(self):
return "BB | Cls: %s, x: %f, y: %f, l: %f, w: %f, yaw: %f" % (
self.cls, self.x, self.y, self.length, self.width, self.yaw)
def rotational_nms(set_boxes, confidences, occ_threshold=0.7, nms_iou_thr=0.5):
""" rotational NMS
set_boxes = size NSeqs list of size NDet lists of tuples. each tuple has the form ((pos, pos), (size, size), angle)
confidences = size NSeqs list of lists containing NDet floats, i.e. one per detection
"""
assert len(set_boxes) == len(confidences) and 0 < occ_threshold < 1 and 0 < nms_iou_thr < 1
if not len(set_boxes):
return []
assert (isinstance(set_boxes[0][0][0][0], float) or isinstance(set_boxes[0][0][0][0], int)) and \
(isinstance(confidences[0][0], float) or isinstance(confidences[0][0], int))
nms_boxes = []
for boxes, confs in zip(set_boxes, confidences):
assert len(boxes) == len(confs)
indices = cv.dnn.NMSBoxesRotated(boxes, confs, occ_threshold, nms_iou_thr)
indices = indices.reshape(len(indices)).tolist()
nms_boxes.append([boxes[i] for i in indices])
return nms_boxes
def generate_bboxes_from_pred(occ, pos, siz, ang, hdg, clf, anchor_dims, occ_threshold=0.5):
""" Generating the bounding boxes based on the regression targets """
# Get only the boxes where occupancy is greater or equal threshold.
real_boxes = np.where(occ >= occ_threshold)
# Get the indices of the occupancy array
coordinates = list(zip(real_boxes[0], real_boxes[1], real_boxes[2]))
# Assign anchor dimensions as original bounding box coordinates which will eventually be changed
# according to the predicted regression targets
anchor_dims = anchor_dims
real_anchors = np.random.rand(len(coordinates), len(anchor_dims[0]))
for i, value in enumerate(real_boxes[2]):
real_anchors[i, ...] = anchor_dims[value]
# Change the anchor boxes based on regression targets, this is the inverse of the operations given in
# createPillarTargets function (src/PointPillars.cpp)
predicted_boxes = []
for i, value in enumerate(coordinates):
real_diag = np.sqrt(np.square(real_anchors[i][0]) + np.square(real_anchors[i][1]))
real_x = value[0] * Parameters.x_step * Parameters.downscaling_factor + Parameters.x_min
real_y = value[1] * Parameters.y_step * Parameters.downscaling_factor + Parameters.y_min
bb_x = pos[value][0] * real_diag + real_x
bb_y = pos[value][1] * real_diag + real_y
bb_z = pos[value][2] * real_anchors[i][2] + real_anchors[i][3]
# print(position[value], real_x, real_y, real_diag)
bb_length = np.exp(siz[value][0]) * real_anchors[i][0]
bb_width = np.exp(siz[value][1]) * real_anchors[i][1]
bb_height = np.exp(siz[value][2]) * real_anchors[i][2]
bb_yaw = -np.arcsin(np.clip(ang[value], -1, 1)) + real_anchors[i][4]
bb_heading = np.round(hdg[value])
bb_cls = np.argmax(clf[value])
bb_conf = occ[value]
predicted_boxes.append(BBox(bb_x, bb_y, bb_z, bb_length, bb_width, bb_height,
bb_yaw, bb_heading, bb_cls, bb_conf))
return predicted_boxes
class GroundTruthGenerator(DataProcessor):
""" Multiprocessing-safe data generator for training, validation or testing, without fancy augmentation """
def __init__(self, data_reader: DataReader, label_files: List[str], calibration_files: List[str] = None,
network_format: bool = False):
super(GroundTruthGenerator, self).__init__()
self.data_reader = data_reader
self.label_files = label_files
self.calibration_files = calibration_files
self.network_format = network_format
def __len__(self):
return len(self.label_files)
def __getitem__(self, file_id: int):
label = self.data_reader.read_label(self.label_files[file_id])
R, t = self.data_reader.read_calibration(self.calibration_files[file_id])
label_transformed = self.transform_labels_into_lidar_coordinates(label, R, t)
if self.network_format:
occupancy, position, size, angle, heading, classification = self.make_ground_truth(label_transformed)
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 [occupancy, position, size, angle, heading, classification]
return label_transformed
def focal_loss_checker(y_true, y_pred, n_occs=-1):
y_true = np.stack(np.where(y_true == 1))
if n_occs == -1:
n_occs = y_true.shape[1]
occ_thr = np.sort(y_pred.flatten())[-n_occs]
y_pred = np.stack(np.where(y_pred >= occ_thr))
p = 0
for gt in range(y_true.shape[1]):
for pr in range(y_pred.shape[1]):
if np.all(y_true[:, gt] == y_pred[:, pr]):
p += 1
break
print("#matched gt: ", p, " #unmatched gt: ", y_true.shape[1] - p, " #unmatched pred: ", y_pred.shape[1] - p,
" occupancy threshold: ", occ_thr)