Spaces:
Running
Running
File size: 15,818 Bytes
f90241e |
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 |
#!/usr/bin/env python3
# Copyright (C) 2024-present Naver Corporation. All rights reserved.
# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
#
# --------------------------------------------------------
# Script to pre-process the arkitscenes dataset.
# Usage:
# python3 datasets_preprocess/preprocess_arkitscenes.py --arkitscenes_dir /path/to/arkitscenes --precomputed_pairs /path/to/arkitscenes_pairs
# --------------------------------------------------------
import os
import json
import os.path as osp
import decimal
import argparse
import math
from bisect import bisect_left
from PIL import Image
import numpy as np
import quaternion
from scipy import interpolate
import cv2
def get_parser():
parser = argparse.ArgumentParser()
parser.add_argument('--arkitscenes_dir', required=True)
parser.add_argument('--precomputed_pairs', required=True)
parser.add_argument('--output_dir', default='data/arkitscenes_processed')
return parser
def value_to_decimal(value, decimal_places):
decimal.getcontext().rounding = decimal.ROUND_HALF_UP # define rounding method
return decimal.Decimal(str(float(value))).quantize(decimal.Decimal('1e-{}'.format(decimal_places)))
def closest(value, sorted_list):
index = bisect_left(sorted_list, value)
if index == 0:
return sorted_list[0]
elif index == len(sorted_list):
return sorted_list[-1]
else:
value_before = sorted_list[index - 1]
value_after = sorted_list[index]
if value_after - value < value - value_before:
return value_after
else:
return value_before
def get_up_vectors(pose_device_to_world):
return np.matmul(pose_device_to_world, np.array([[0.0], [-1.0], [0.0], [0.0]]))
def get_right_vectors(pose_device_to_world):
return np.matmul(pose_device_to_world, np.array([[1.0], [0.0], [0.0], [0.0]]))
def read_traj(traj_path):
quaternions = []
poses = []
timestamps = []
poses_p_to_w = []
with open(traj_path) as f:
traj_lines = f.readlines()
for line in traj_lines:
tokens = line.split()
assert len(tokens) == 7
traj_timestamp = float(tokens[0])
timestamps_decimal_value = value_to_decimal(traj_timestamp, 3)
timestamps.append(float(timestamps_decimal_value)) # for spline interpolation
angle_axis = [float(tokens[1]), float(tokens[2]), float(tokens[3])]
r_w_to_p, _ = cv2.Rodrigues(np.asarray(angle_axis))
t_w_to_p = np.asarray([float(tokens[4]), float(tokens[5]), float(tokens[6])])
pose_w_to_p = np.eye(4)
pose_w_to_p[:3, :3] = r_w_to_p
pose_w_to_p[:3, 3] = t_w_to_p
pose_p_to_w = np.linalg.inv(pose_w_to_p)
r_p_to_w_as_quat = quaternion.from_rotation_matrix(pose_p_to_w[:3, :3])
t_p_to_w = pose_p_to_w[:3, 3]
poses_p_to_w.append(pose_p_to_w)
poses.append(t_p_to_w)
quaternions.append(r_p_to_w_as_quat)
return timestamps, poses, quaternions, poses_p_to_w
def main(rootdir, pairsdir, outdir):
os.makedirs(outdir, exist_ok=True)
subdirs = ['Test', 'Training']
for subdir in subdirs:
if not osp.isdir(osp.join(rootdir, subdir)):
continue
# STEP 1: list all scenes
outsubdir = osp.join(outdir, subdir)
os.makedirs(outsubdir, exist_ok=True)
listfile = osp.join(pairsdir, subdir, 'scene_list.json')
with open(listfile, 'r') as f:
scene_dirs = json.load(f)
valid_scenes = []
for scene_subdir in scene_dirs:
out_scene_subdir = osp.join(outsubdir, scene_subdir)
os.makedirs(out_scene_subdir, exist_ok=True)
scene_dir = osp.join(rootdir, subdir, scene_subdir)
depth_dir = osp.join(scene_dir, 'lowres_depth')
rgb_dir = osp.join(scene_dir, 'vga_wide')
intrinsics_dir = osp.join(scene_dir, 'vga_wide_intrinsics')
traj_path = osp.join(scene_dir, 'lowres_wide.traj')
# STEP 2: read selected_pairs.npz
selected_pairs_path = osp.join(pairsdir, subdir, scene_subdir, 'selected_pairs.npz')
selected_npz = np.load(selected_pairs_path)
selection, pairs = selected_npz['selection'], selected_npz['pairs']
selected_sky_direction_scene = str(selected_npz['sky_direction_scene'][0])
if len(selection) == 0 or len(pairs) == 0:
# not a valid scene
continue
valid_scenes.append(scene_subdir)
# STEP 3: parse the scene and export the list of valid (K, pose, rgb, depth) and convert images
scene_metadata_path = osp.join(out_scene_subdir, 'scene_metadata.npz')
if osp.isfile(scene_metadata_path):
continue
else:
print(f'parsing {scene_subdir}')
# loads traj
timestamps, poses, quaternions, poses_cam_to_world = read_traj(traj_path)
poses = np.array(poses)
quaternions = np.array(quaternions, dtype=np.quaternion)
quaternions = quaternion.unflip_rotors(quaternions)
timestamps = np.array(timestamps)
selected_images = [(basename, basename.split(".png")[0].split("_")[1]) for basename in selection]
timestamps_selected = [float(frame_id) for _, frame_id in selected_images]
sky_direction_scene, trajectories, intrinsics, images = convert_scene_metadata(scene_subdir,
intrinsics_dir,
timestamps,
quaternions,
poses,
poses_cam_to_world,
selected_images,
timestamps_selected)
assert selected_sky_direction_scene == sky_direction_scene
os.makedirs(os.path.join(out_scene_subdir, 'vga_wide'), exist_ok=True)
os.makedirs(os.path.join(out_scene_subdir, 'lowres_depth'), exist_ok=True)
assert isinstance(sky_direction_scene, str)
for basename in images:
img_out = os.path.join(out_scene_subdir, 'vga_wide', basename.replace('.png', '.jpg'))
depth_out = os.path.join(out_scene_subdir, 'lowres_depth', basename)
if osp.isfile(img_out) and osp.isfile(depth_out):
continue
vga_wide_path = osp.join(rgb_dir, basename)
depth_path = osp.join(depth_dir, basename)
img = Image.open(vga_wide_path)
depth = cv2.imread(depth_path, cv2.IMREAD_UNCHANGED)
# rotate the image
if sky_direction_scene == 'RIGHT':
try:
img = img.transpose(Image.Transpose.ROTATE_90)
except Exception:
img = img.transpose(Image.ROTATE_90)
depth = cv2.rotate(depth, cv2.ROTATE_90_COUNTERCLOCKWISE)
elif sky_direction_scene == 'LEFT':
try:
img = img.transpose(Image.Transpose.ROTATE_270)
except Exception:
img = img.transpose(Image.ROTATE_270)
depth = cv2.rotate(depth, cv2.ROTATE_90_CLOCKWISE)
elif sky_direction_scene == 'DOWN':
try:
img = img.transpose(Image.Transpose.ROTATE_180)
except Exception:
img = img.transpose(Image.ROTATE_180)
depth = cv2.rotate(depth, cv2.ROTATE_180)
W, H = img.size
if not osp.isfile(img_out):
img.save(img_out)
depth = cv2.resize(depth, (W, H), interpolation=cv2.INTER_NEAREST_EXACT)
if not osp.isfile(depth_out): # avoid destroying the base dataset when you mess up the paths
cv2.imwrite(depth_out, depth)
# save at the end
np.savez(scene_metadata_path,
trajectories=trajectories,
intrinsics=intrinsics,
images=images,
pairs=pairs)
outlistfile = osp.join(outsubdir, 'scene_list.json')
with open(outlistfile, 'w') as f:
json.dump(valid_scenes, f)
# STEP 5: concat all scene_metadata.npz into a single file
scene_data = {}
for scene_subdir in valid_scenes:
scene_metadata_path = osp.join(outsubdir, scene_subdir, 'scene_metadata.npz')
with np.load(scene_metadata_path) as data:
trajectories = data['trajectories']
intrinsics = data['intrinsics']
images = data['images']
pairs = data['pairs']
scene_data[scene_subdir] = {'trajectories': trajectories,
'intrinsics': intrinsics,
'images': images,
'pairs': pairs}
offset = 0
counts = []
scenes = []
sceneids = []
images = []
intrinsics = []
trajectories = []
pairs = []
for scene_idx, (scene_subdir, data) in enumerate(scene_data.items()):
num_imgs = data['images'].shape[0]
img_pairs = data['pairs']
scenes.append(scene_subdir)
sceneids.extend([scene_idx] * num_imgs)
images.append(data['images'])
K = np.expand_dims(np.eye(3), 0).repeat(num_imgs, 0)
K[:, 0, 0] = [fx for _, _, fx, _, _, _ in data['intrinsics']]
K[:, 1, 1] = [fy for _, _, _, fy, _, _ in data['intrinsics']]
K[:, 0, 2] = [hw for _, _, _, _, hw, _ in data['intrinsics']]
K[:, 1, 2] = [hh for _, _, _, _, _, hh in data['intrinsics']]
intrinsics.append(K)
trajectories.append(data['trajectories'])
# offset pairs
img_pairs[:, 0:2] += offset
pairs.append(img_pairs)
counts.append(offset)
offset += num_imgs
images = np.concatenate(images, axis=0)
intrinsics = np.concatenate(intrinsics, axis=0)
trajectories = np.concatenate(trajectories, axis=0)
pairs = np.concatenate(pairs, axis=0)
np.savez(osp.join(outsubdir, 'all_metadata.npz'),
counts=counts,
scenes=scenes,
sceneids=sceneids,
images=images,
intrinsics=intrinsics,
trajectories=trajectories,
pairs=pairs)
def convert_scene_metadata(scene_subdir, intrinsics_dir,
timestamps, quaternions, poses, poses_cam_to_world,
selected_images, timestamps_selected):
# find scene orientation
sky_direction_scene, rotated_to_cam = find_scene_orientation(poses_cam_to_world)
# find/compute pose for selected timestamps
# most images have a valid timestamp / exact pose associated
timestamps_selected = np.array(timestamps_selected)
spline = interpolate.interp1d(timestamps, poses, kind='linear', axis=0)
interpolated_rotations = quaternion.squad(quaternions, timestamps, timestamps_selected)
interpolated_positions = spline(timestamps_selected)
trajectories = []
intrinsics = []
images = []
for i, (basename, frame_id) in enumerate(selected_images):
intrinsic_fn = osp.join(intrinsics_dir, f"{scene_subdir}_{frame_id}.pincam")
if not osp.exists(intrinsic_fn):
intrinsic_fn = osp.join(intrinsics_dir, f"{scene_subdir}_{float(frame_id) - 0.001:.3f}.pincam")
if not osp.exists(intrinsic_fn):
intrinsic_fn = osp.join(intrinsics_dir, f"{scene_subdir}_{float(frame_id) + 0.001:.3f}.pincam")
assert osp.exists(intrinsic_fn)
w, h, fx, fy, hw, hh = np.loadtxt(intrinsic_fn) # PINHOLE
pose = np.eye(4)
pose[:3, :3] = quaternion.as_rotation_matrix(interpolated_rotations[i])
pose[:3, 3] = interpolated_positions[i]
images.append(basename)
if sky_direction_scene == 'RIGHT' or sky_direction_scene == 'LEFT':
intrinsics.append([h, w, fy, fx, hh, hw]) # swapped intrinsics
else:
intrinsics.append([w, h, fx, fy, hw, hh])
trajectories.append(pose @ rotated_to_cam) # pose_cam_to_world @ rotated_to_cam = rotated(cam) to world
return sky_direction_scene, trajectories, intrinsics, images
def find_scene_orientation(poses_cam_to_world):
if len(poses_cam_to_world) > 0:
up_vector = sum(get_up_vectors(p) for p in poses_cam_to_world) / len(poses_cam_to_world)
right_vector = sum(get_right_vectors(p) for p in poses_cam_to_world) / len(poses_cam_to_world)
up_world = np.array([[0.0], [0.0], [1.0], [0.0]])
else:
up_vector = np.array([[0.0], [-1.0], [0.0], [0.0]])
right_vector = np.array([[1.0], [0.0], [0.0], [0.0]])
up_world = np.array([[0.0], [0.0], [1.0], [0.0]])
# value between 0, 180
device_up_to_world_up_angle = np.arccos(np.clip(np.dot(np.transpose(up_world),
up_vector), -1.0, 1.0)).item() * 180.0 / np.pi
device_right_to_world_up_angle = np.arccos(np.clip(np.dot(np.transpose(up_world),
right_vector), -1.0, 1.0)).item() * 180.0 / np.pi
up_closest_to_90 = abs(device_up_to_world_up_angle - 90.0) < abs(device_right_to_world_up_angle - 90.0)
if up_closest_to_90:
assert abs(device_up_to_world_up_angle - 90.0) < 45.0
# LEFT
if device_right_to_world_up_angle > 90.0:
sky_direction_scene = 'LEFT'
cam_to_rotated_q = quaternion.from_rotation_vector([0.0, 0.0, math.pi / 2.0])
else:
# note that in metadata.csv RIGHT does not exist, but again it's not accurate...
# well, turns out there are scenes oriented like this
# for example Training/41124801
sky_direction_scene = 'RIGHT'
cam_to_rotated_q = quaternion.from_rotation_vector([0.0, 0.0, -math.pi / 2.0])
else:
# right is close to 90
assert abs(device_right_to_world_up_angle - 90.0) < 45.0
if device_up_to_world_up_angle > 90.0:
sky_direction_scene = 'DOWN'
cam_to_rotated_q = quaternion.from_rotation_vector([0.0, 0.0, math.pi])
else:
sky_direction_scene = 'UP'
cam_to_rotated_q = quaternion.quaternion(1, 0, 0, 0)
cam_to_rotated = np.eye(4)
cam_to_rotated[:3, :3] = quaternion.as_rotation_matrix(cam_to_rotated_q)
rotated_to_cam = np.linalg.inv(cam_to_rotated)
return sky_direction_scene, rotated_to_cam
if __name__ == '__main__':
parser = get_parser()
args = parser.parse_args()
main(args.arkitscenes_dir, args.precomputed_pairs, args.output_dir)
|