forked from commaai/speedchallenge
-
Notifications
You must be signed in to change notification settings - Fork 1
/
test_pose_augmentation.py
161 lines (109 loc) · 5.77 KB
/
test_pose_augmentation.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
import numpy as np
import cv2
from tensorflow.python.training.tracking import base
from extract import create_records, write_records
import glob
from multiprocessing import Pool
import camera
import orientation
import tensorflow as tf
import keras_resnet
import depth_and_motion_net
import random
def create_model():
inputs = tf.keras.Input(shape=(128, 416, 6), name='frames')
# encoder
conv5 = keras_resnet.resnet18_encoder(inputs)
# conv5 = keras_regnet.regnety_400mf(inputs)
# more downsampling
conv6 = keras_resnet.res_block_first(conv5, 512, stride=2)
conv6 = keras_resnet.res_block(conv6, 512)
conv7 = keras_resnet.res_block_first(conv6, 512, stride=2)
conv7 = keras_resnet.res_block(conv7, 512)
bottleneck = tf.keras.layers.Lambda(lambda x: tf.reduce_mean(x, [1, 2]))(conv7)
# bottleneck = tf.keras.layers.Dropout(0.2)(bottleneck)
rotation = tf.keras.layers.Dense(3, name='rot_fc')(bottleneck)
translation = tf.keras.layers.Dense(3, name='trans_fc')(bottleneck)
rotation = depth_and_motion_net.Scale(0.001)(rotation)
translation = depth_and_motion_net.Scale(0.001)(translation)
# translation, rotation = depth_and_motion_net.depth_and_motion_net_fc_no_mean(inputs)
pose = tf.keras.layers.Concatenate(axis=1, name='pose')([translation, rotation])
# speed prediction
speed = tf.keras.layers.Lambda(lambda x: tf.expand_dims(20 * tf.norm(x[:, :3], axis=1), -1), name='speed')(pose)
model = tf.keras.Model(inputs=inputs, outputs=[pose, speed])
model.compile(optimizer=tf.keras.optimizers.Adam(1e-4), loss={'pose': 'mse', 'speed': 'mse'}, loss_weights={'pose': 1.0, 'speed': 0.0})
latest = tf.train.latest_checkpoint("checkpoints/07_07_21-11_07")
model.load_weights(latest).expect_partial()
return model
def resize_crop_and_standardize(frame):
frame = cv2.resize(frame, (640, 480), interpolation=cv2.INTER_AREA)
frame = frame[142:339, :]
frame = cv2.resize(frame, (416, 128), interpolation=cv2.INTER_AREA)
frame = tf.image.convert_image_dtype(frame, tf.float32)
frame = tf.image.per_image_standardization(frame)
return frame
# takes in 2 ecef quaternions and returns the euler angle between them
def relative_orientation(ecef_quat_1, ecef_quat_2):
ecef_mat_1 = orientation.quat2rot(ecef_quat_1)
ecef_mat_2 = orientation.quat2rot(ecef_quat_2)
relative_mat = np.matmul(ecef_mat_2, np.transpose(ecef_mat_1))
return orientation.rot2euler(relative_mat)
def extract_segment(path):
# radar_returns = np.load(path + '/processed_log/CAN/radar/value')
# radar_time_stamps = np.load(path + '/processed_log/CAN/radar/t')
# print(radar_returns.shape)
# for rr, t in zip(radar_returns[:10], radar_time_stamps[:10]):
# print(t, rr)
# return
model = create_model()
frame_velocities = np.linalg.norm(np.load(path + '/global_pose/frame_velocities'),axis=1)
frame_velocities = list(map(lambda x: x.item(), frame_velocities))
cam = cv2.VideoCapture(path + "/video.hevc")
frames = []
for i in range(10):
ret, frame = cam.read()
frames.append(frame)
frame_positions = np.load(path + '/global_pose/frame_positions')
frame_orientations = np.load(path + '/global_pose/frame_orientations')
for i in range(len(frame_positions[:-3])):
plus_one_position = camera.device_from_ecef(frame_positions[i], frame_orientations[i], frame_positions[i+1])
plus_one_orientation = relative_orientation(frame_orientations[i], frame_orientations[i+1])
base_frame = frames[i]
base_frame_standard = resize_crop_and_standardize(base_frame)
# cv2.imwrite("test1.jpg", base_frame)
plus_one_real = frames[i+1]
# plus_one_real = resize_crop_and_standardize(plus_one_real)
# cv2.imwrite("test2.jpg", plus_one_real)
# input = tf.concat((base_frame, plus_one_real), axis=2)
# input = tf.expand_dims(input, axis=0)
# pose, speed = model(input)
# print(pose)
# plp_view = camera.view_frame_from_device_frame.dot(np.array(plus_one_position))
# print(plp_view)
# plus_one_fake = camera.transform_img(np.array(base_frame), augment_trans=plp_view)
# plus_one_fake = resize_crop_and_standardize(plus_one_fake)
# input = tf.concat((base_frame, plus_one_fake), axis=2)
# input = tf.expand_dims(input, axis=0)
# pose, speed = model(input)
# print(pose)
# xplus = camera.transform_img(np.array(base_frame), augment_trans=np.array([0.1, 0, 0]))
# cv2.imwrite("plus_one_real.jpg", plus_one_real)
# # cv2.imwrite("plus_one_fake.jpg", plus_one_fake)
# break
xplus = camera.transform_img(np.array(plus_one_real), augment_trans=camera.view_frame_from_device_frame.dot(np.array((.5, 0.0, 0.0))))
xplus_standard = resize_crop_and_standardize(xplus)
input = tf.concat((base_frame_standard, xplus_standard), axis=2)
input = tf.expand_dims(input, axis=0)
pose, speed = model(input)
print(plus_one_position)
print(pose[:, :3])
# cv2.imwrite("plus_one_fake.jpg", xplus_standard)
# xminus = camera.transform_img(np.array(base_frame), augment_trans=np.array([-0.1, 0, 0]))
# yplus = camera.transform_img(np.array(base_frame), augment_trans=np.array([0, 0.1, 0]))
# yminus = camera.transform_img(np.array(base_frame), augment_trans=np.array([0, -0.1, 0]))
# zplus = camera.transform_img(np.array(base_frame), augment_trans=np.array([0, 0, 0.1]))
# zminus = camera.transform_img(np.array(base_frame), augment_trans=np.array([0, 0, -0.1]))
# print(plus_one_position)
# print(plus_one_orientation)
break
extract_segment("/mnt/e/commaai/comma2k19/Chunk_1/b0c9d2329ad1606b_2018-07-27--06-03-57/11")