File size: 13,192 Bytes
f348660 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 | import torch
import torch.utils.data
from torch.utils.data import Dataset
import numpy as np
import data.data_utils as data_utils
import random
import os
from PIL import Image
from data.data_utils import load_depth
def random_sample(T):
'''
Arg(s):
T : numpy[float32]
C x N array
Returns:
numpy[float32] : random sample from T
'''
index = np.random.randint(0, T.shape[0])
return T[index, :]
def random_crop(inputs, shape, crop_type=['none']):
'''
Apply crop to inputs e.g. images, depth
Arg(s):
inputs : list[numpy[float32]]
list of numpy arrays e.g. images, depth, and validity maps
shape : list[int]
shape (height, width) to crop inputs
crop_type : str
none, horizontal, vertical, anchored, top, bottom, left, right, center
Return:
list[numpy[float32]] : list of cropped inputs
'''
n_height, n_width = shape
_, o_height, o_width = inputs[0].shape
# Get delta of crop and original height and width
d_height = o_height - n_height
d_width = o_width - n_width
# By default, perform center crop
y_start = d_height // 2
x_start = d_width // 2
# If left alignment, then set starting height to 0
if 'left' in crop_type:
x_start = 0
# If right alignment, then set starting height to right most position
elif 'right' in crop_type:
x_start = d_width
elif 'horizontal' in crop_type:
# Select from one of the pre-defined anchored locations
if 'anchored' in crop_type:
# Create anchor positions
crop_anchors = [
0.0, 0.50, 1.0
]
widths = [
anchor * d_width for anchor in crop_anchors
]
x_start = int(widths[np.random.randint(low=0, high=len(widths))])
# Randomly select a crop location
else:
x_start = np.random.randint(low=0, high=d_width)
# If top alignment, then set starting height to 0
if 'top' in crop_type:
y_start = 0
# If bottom alignment, then set starting height to lowest position
elif 'bottom' in crop_type:
y_start = d_height
elif 'vertical' in crop_type and np.random.rand() <= 0.30:
# Select from one of the pre-defined anchored locations
if 'anchored' in crop_type:
# Create anchor positions
crop_anchors = [
0.0, 0.50, 1.0
]
heights = [
anchor * d_height for anchor in crop_anchors
]
y_start = int(heights[np.random.randint(low=0, high=len(heights))])
# Randomly select a crop location
else:
y_start = np.random.randint(low=0, high=d_height)
elif 'center' in crop_type:
pass
# Crop each input into (n_height, n_width)
y_end = y_start + n_height
x_end = x_start + n_width
outputs = [
T[:, y_start:y_end, x_start:x_end] for T in inputs
]
return outputs
class RCNetTrainingDataset(torch.utils.data.Dataset):
'''
Dataset for fetching:
(1) image
(2) radar point
(3) ground truth
(4) bounding boxes for the points
(5) image crops for summary part of the code
Arg(s):
image_paths : list[str]
paths to images
radar_paths : list[str]
paths to radar points
ground_truth_paths : list[str]
paths to ground truth depth maps
crop_width : int
width of crop centered at the radar point
total_points_sampled: int
total number of points sampled from the total radar points available. Repeats the same points multiple times if total points in the frame is less than total sampled points
sample_probability_of_lidar: int
randomly sample lidar with this probability and add noise to it instead of using radar points
min_radar_depth_m: float
minimum depth accepted for synthetic radar sampling
max_radar_depth_m: float
maximum depth accepted for synthetic radar sampling
'''
def __init__(self,
image_paths,
radar_paths,
ground_truth_paths,
patch_size,
total_points_sampled,
sample_probability_of_lidar,
min_radar_depth_m=0.05,
max_radar_depth_m=11.2):
self.n_sample = len(image_paths)
assert self.n_sample == len(ground_truth_paths)
assert self.n_sample == len(radar_paths)
self.image_paths = image_paths
self.radar_paths = radar_paths
self.ground_truth_paths = ground_truth_paths
self.patch_size = patch_size
self.pad_size_x = patch_size[1] // 2
self.padding = ((0, 0), (0, 0), (self.pad_size_x, self.pad_size_x))
self.data_format = 'CHW'
self.total_points_sampled = total_points_sampled
self.sample_probability_of_lidar = sample_probability_of_lidar
self.min_radar_depth_m = min_radar_depth_m
self.max_radar_depth_m = max_radar_depth_m
def __getitem__(self, index):
# Load image
image = data_utils.load_image(
self.image_paths[index],
normalize=False,
data_format=self.data_format)
height, width = image.shape[1:]
if height == 720: # ZJU dataset
image = image[:, 720 // 3: 720 // 4 * 3, :]
image = np.pad(
image,
pad_width=self.padding,
mode='edge')
# Load radar points N x 3
radar_points = np.load(self.radar_paths[index])
if height == 720:
radar_points = radar_points[radar_points[:, 1] < 720 // 4 * 3]
radar_points[:, 1] = radar_points[:, 1] - 720 // 3
radar_points = radar_points[radar_points[:, 1] >= 0]
if radar_points.ndim == 1:
# Only one point (,3), expand to 1 x 3
radar_points = np.expand_dims(radar_points, axis=0)
# Store bounding boxes for all radar points
bounding_boxes_list = []
# randomly sample radar points to output
if radar_points.shape[0] <= self.total_points_sampled:
radar_points = np.repeat(radar_points, 100, axis=0)
random_idx = np.random.randint(radar_points.shape[0], size=self.total_points_sampled)
radar_points = radar_points[random_idx, :]
# Load ground truth depth
ground_truth = data_utils.load_depth(
self.ground_truth_paths[index],
data_format=self.data_format)
if height == 720:
ground_truth = ground_truth[:, 720 // 3: 720 // 4 * 3]
if random.random() < self.sample_probability_of_lidar:
ground_truth_for_sampling = np.copy(ground_truth)
ground_truth_for_sampling = ground_truth_for_sampling.squeeze()
valid_lidar = np.isfinite(ground_truth_for_sampling)
valid_lidar &= ground_truth_for_sampling >= self.min_radar_depth_m
valid_lidar &= ground_truth_for_sampling <= self.max_radar_depth_m
idx_lidar_samples = np.where(valid_lidar)
n_lidar_samples = len(idx_lidar_samples[0])
if n_lidar_samples > 0:
# Keep the fixed point count required by RC-Net. Replacement
# handles frames with fewer valid GT pixels than requested.
if n_lidar_samples >= self.total_points_sampled:
random_indices = random.sample(
range(n_lidar_samples), self.total_points_sampled
)
else:
random_indices = np.random.choice(
n_lidar_samples,
size=self.total_points_sampled,
replace=True,
)
points_x = idx_lidar_samples[1][random_indices]
points_y = idx_lidar_samples[0][random_indices]
points_z = ground_truth_for_sampling[points_y, points_x]
noise_for_fake_radar_x = np.random.normal(0, 25, radar_points.shape[0])
noise_for_fake_radar_z = np.random.uniform(low=0.0, high=0.4, size=radar_points.shape[0])
fake_radar_points = np.copy(radar_points)
fake_radar_points[:, 0] = points_x + noise_for_fake_radar_x
fake_radar_points[:, 0] = np.clip(fake_radar_points[:, 0], 0, ground_truth_for_sampling.shape[1])
fake_radar_points[:, 2] = points_z + noise_for_fake_radar_z
# we keep the y as the same it is since it is erroneous
# convert x and y indices back to int after adding noise
fake_radar_points[:, 0] = fake_radar_points[:, 0].astype(int)
fake_radar_points[:, 1] = fake_radar_points[:, 1].astype(int)
radar_points = np.copy(fake_radar_points)
# get the shifted radar points after padding
for radar_point_idx in range(0, radar_points.shape[0]):
# Set radar point to the center of the patch
radar_points[radar_point_idx, 0] = radar_points[radar_point_idx, 0] + self.pad_size_x
bounding_box = [0, 0, 0, 0]
bounding_box[0] = radar_points[radar_point_idx, 0] - self.pad_size_x
bounding_box[1] = 0
bounding_box[2] = radar_points[radar_point_idx, 0] + self.pad_size_x
bounding_box[3] = self.patch_size[0]
bounding_boxes_list.append(np.asarray(bounding_box))
ground_truth = np.pad(
ground_truth,
pad_width=self.padding,
mode='constant',
constant_values=0)
ground_truth_crops = []
# Crop image and ground truth
for radar_point_idx in range(0, radar_points.shape[0]):
start_x = int(radar_points[radar_point_idx, 0] - self.pad_size_x)
end_x = int(radar_points[radar_point_idx, 0] + self.pad_size_x)
start_y = image.shape[-2] - self.patch_size[0]
ground_truth_cropped = ground_truth[:, start_y:, start_x:end_x]
ground_truth_crops.append(ground_truth_cropped)
image = image[:, start_y:, ...]
ground_truth = np.asarray(ground_truth_crops)
# Convert to float32
image, radar_points, ground_truth = [
T.astype(np.float32)
for T in [image, radar_points, ground_truth]
]
bounding_boxes_list = [T.astype(np.float32) for T in bounding_boxes_list]
bounding_boxes_list = np.stack(bounding_boxes_list, axis=0)
return image, radar_points, bounding_boxes_list, ground_truth
def __len__(self):
return self.n_sample
class RCNetInferenceDataset(torch.utils.data.Dataset):
'''
Dataset for fetching:
(1) image
(2) radar points
(3) ground truth (if available)
Arg(s):
image_paths : list[str]
paths to images
radar_paths : list[str]
paths to radar points
ground_truth_paths : list[str]
paths to ground truth paths
'''
def __init__(self, image_paths, radar_paths, ground_truth_paths=None):
self.n_sample = len(image_paths)
assert self.n_sample == len(radar_paths)
self.image_paths = image_paths
self.radar_paths = radar_paths
if ground_truth_paths is not None and None not in ground_truth_paths:
assert self.n_sample == len(ground_truth_paths)
self.ground_truth_available = True
else:
self.ground_truth_available = False
self.ground_truth_paths = ground_truth_paths
self.data_format = 'CHW'
def __getitem__(self, index):
# Load image
image = data_utils.load_image(
self.image_paths[index],
normalize=False,
data_format=self.data_format)
height, width = image.shape[1:]
if height == 720: # ZJU dataset
image = image[:, 720 // 3: 720 // 4 * 3, :]
# Load radar points N x 3
radar_points = np.load(self.radar_paths[index])
if height == 720:
radar_points = radar_points[radar_points[:, 1] < 720 // 4 * 3]
radar_points[:, 1] = radar_points[:, 1] - 720 // 3
radar_points = radar_points[radar_points[:, 1] >= 0]
if radar_points.ndim == 1:
# Expand to 1 x 3
radar_points = np.expand_dims(radar_points, axis=0)
inputs = [image, radar_points]
if self.ground_truth_available:
# Load ground truth depth
ground_truth = data_utils.load_depth(
self.ground_truth_paths[index],
data_format=self.data_format)
if height == 720:
ground_truth = ground_truth[:, 720 // 3: 720 // 4 * 3]
inputs.append(ground_truth)
# Convert to float32
inputs = [
T.astype(np.float32)
for T in inputs
]
return inputs
def __len__(self):
return self.n_sample
|