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
|
"""
An example for running incremental SfM on 360 spherical panorama images.
"""
import argparse
import os
from collections.abc import Sequence
from concurrent.futures import ThreadPoolExecutor, as_completed
from dataclasses import dataclass
from pathlib import Path
from threading import Lock
import cv2
import numpy as np
import PIL.ExifTags
import PIL.Image
from scipy.spatial.transform import Rotation
from tqdm import tqdm
import pycolmap
from pycolmap import logging
@dataclass
class PanoRenderOptions:
num_steps_yaw: int
pitches_deg: Sequence[float]
hfov_deg: float
vfov_deg: float
PANO_RENDER_OPTIONS: dict[str, PanoRenderOptions] = {
"overlapping": PanoRenderOptions(
num_steps_yaw=4,
pitches_deg=(-35.0, 0.0, 35.0),
hfov_deg=90.0,
vfov_deg=90.0,
),
# Cubemap without top and bottom images.
"non-overlapping": PanoRenderOptions(
num_steps_yaw=4,
pitches_deg=(0.0,),
hfov_deg=90.0,
vfov_deg=90.0,
),
}
def create_virtual_camera(
pano_width: int,
pano_height: int,
hfov_deg: float,
vfov_deg: float,
) -> pycolmap.Camera:
"""Create a virtual perspective camera."""
image_width = int(pano_width * hfov_deg / 360)
image_height = int(pano_height * vfov_deg / 180)
focal = image_width / (2 * np.tan(np.deg2rad(hfov_deg) / 2))
return pycolmap.Camera.create(
0, "SIMPLE_PINHOLE", focal, image_width, image_height
)
def get_virtual_camera_rays(camera: pycolmap.Camera) -> np.ndarray:
size = (camera.width, camera.height)
x, y = np.indices(size).astype(np.float32)
xy = np.column_stack([x.ravel(), y.ravel()])
# The center of the upper left most pixel has coordinate (0.5, 0.5)
xy += 0.5
xy_norm = camera.cam_from_img(xy)
rays = np.concatenate([xy_norm, np.ones_like(xy_norm[:, :1])], -1)
rays /= np.linalg.norm(rays, axis=-1, keepdims=True)
return rays
def spherical_img_from_cam(image_size, rays_in_cam: np.ndarray) -> np.ndarray:
"""Project rays into a 360 panorama (spherical) image."""
if image_size[0] != image_size[1] * 2:
raise ValueError("Only 360° panoramas are supported.")
if rays_in_cam.ndim != 2 or rays_in_cam.shape[1] != 3:
raise ValueError(f"{rays_in_cam.shape=} but expected (N,3).")
r = rays_in_cam.T
yaw = np.arctan2(r[0], r[2])
pitch = -np.arctan2(r[1], np.linalg.norm(r[[0, 2]], axis=0))
u = (1 + yaw / np.pi) / 2
v = (1 - pitch * 2 / np.pi) / 2
return np.stack([u, v], -1) * image_size
def get_virtual_rotations(
num_steps_yaw: int, pitches_deg: Sequence[float]
) -> Sequence[np.ndarray]:
"""Get the relative rotations of the virtual cameras w.r.t. the panorama."""
# Assuming that the panos are approximately upright.
cams_from_pano_r = []
yaws = np.linspace(0, 360, num_steps_yaw, endpoint=False)
for pitch_deg in pitches_deg:
yaw_offset = (360 / num_steps_yaw / 2) if pitch_deg > 0 else 0
for yaw_deg in yaws + yaw_offset:
cam_from_pano_r = Rotation.from_euler(
"XY", [-pitch_deg, -yaw_deg], degrees=True
).as_matrix()
cams_from_pano_r.append(cam_from_pano_r)
return cams_from_pano_r
def create_pano_rig_config(
cams_from_pano_rotation: Sequence[np.ndarray], ref_idx: int = 0
) -> pycolmap.RigConfig:
"""Create a RigConfig for the given virtual rotations."""
rig_cameras = []
for idx, cam_from_pano_rotation in enumerate(cams_from_pano_rotation):
if idx == ref_idx:
cam_from_rig = None
else:
cam_from_ref_rotation = (
cam_from_pano_rotation @ cams_from_pano_rotation[ref_idx].T
)
cam_from_rig = pycolmap.Rigid3d(
pycolmap.Rotation3d(cam_from_ref_rotation), np.zeros(3)
)
rig_cameras.append(
pycolmap.RigConfigCamera(
ref_sensor=idx == ref_idx,
image_prefix=f"pano_camera{idx}/",
cam_from_rig=cam_from_rig,
)
)
return pycolmap.RigConfig(cameras=rig_cameras)
class PanoProcessor:
def __init__(
self,
pano_image_dir: Path,
output_image_dir: Path,
mask_dir: Path,
render_options: PanoRenderOptions,
):
self.render_options = render_options
self.pano_image_dir = pano_image_dir
self.output_image_dir = output_image_dir
self.mask_dir = mask_dir
self.cams_from_pano_rotation = get_virtual_rotations(
num_steps_yaw=render_options.num_steps_yaw,
pitches_deg=render_options.pitches_deg,
)
self.rig_config = create_pano_rig_config(self.cams_from_pano_rotation)
# We assign each pano pixel to the virtual camera
# with the closest camera center.
self.cam_centers_in_pano = np.einsum(
"nij,i->nj", self.cams_from_pano_rotation, [0, 0, 1]
)
self._lock = Lock()
# These are initialized on the first pano image
# to avoid recomputing the rays for each pano image.
self._camera = None
self._pano_size = None
self._rays_in_cam = None
def process(self, pano_name: str):
pano_path = self.pano_image_dir / pano_name
try:
pano_image = PIL.Image.open(pano_path)
except PIL.Image.UnidentifiedImageError:
logging.info(f"Skipping file {pano_path} as it cannot be read.")
return
pano_exif = pano_image.getexif()
pano_image = np.asarray(pano_image)
gpsonly_exif = PIL.Image.Exif()
gpsonly_exif[PIL.ExifTags.IFD.GPSInfo] = pano_exif.get_ifd(
PIL.ExifTags.IFD.GPSInfo
)
pano_height, pano_width, *_ = pano_image.shape
if pano_width != pano_height * 2:
raise ValueError("Only 360° panoramas are supported.")
with self._lock:
if self._camera is None: # First image, precompute rays once.
self._camera = create_virtual_camera(
pano_width=pano_width,
pano_height=pano_height,
hfov_deg=self.render_options.hfov_deg,
vfov_deg=self.render_options.vfov_deg,
)
for rig_camera in self.rig_config.cameras:
rig_camera.camera = self._camera
self._pano_size = (pano_width, pano_height)
self._rays_in_cam = get_virtual_camera_rays(self._camera)
else: # Later images, verify consistent panoramas.
if (pano_width, pano_height) != self._pano_size:
raise ValueError(
"Panoramas of different sizes are not supported."
)
for cam_idx, cam_from_pano_r in enumerate(self.cams_from_pano_rotation):
rays_in_pano = self._rays_in_cam @ cam_from_pano_r
xy_in_pano = spherical_img_from_cam(self._pano_size, rays_in_pano)
xy_in_pano = xy_in_pano.reshape(
self._camera.width, self._camera.height, 2
).astype(np.float32)
xy_in_pano -= 0.5 # COLMAP to OpenCV pixel origin.
image = cv2.remap(
pano_image,
*np.moveaxis(xy_in_pano, [0, 1, 2], [2, 1, 0]),
cv2.INTER_LINEAR,
borderMode=cv2.BORDER_WRAP,
)
# We define a mask such that each pixel of the panorama has its
# features extracted only in a single virtual camera.
closest_camera = np.argmax(
rays_in_pano @ self.cam_centers_in_pano.T, -1
)
mask = (
((closest_camera == cam_idx) * 255)
.astype(np.uint8)
.reshape(self._camera.width, self._camera.height)
.transpose()
)
image_name = (
self.rig_config.cameras[cam_idx].image_prefix + pano_name
)
mask_name = f"{image_name}.png"
image_path = self.output_image_dir / image_name
image_path.parent.mkdir(exist_ok=True, parents=True)
PIL.Image.fromarray(image).save(image_path, exif=gpsonly_exif)
mask_path = self.mask_dir / mask_name
mask_path.parent.mkdir(exist_ok=True, parents=True)
if not pycolmap.Bitmap.from_array(mask).write(mask_path):
raise RuntimeError(f"Cannot write {mask_path}")
def render_perspective_images(
pano_image_names: Sequence[str],
pano_image_dir: Path,
output_image_dir: Path,
mask_dir: Path,
render_options: PanoRenderOptions,
) -> pycolmap.RigConfig:
processor = PanoProcessor(
pano_image_dir, output_image_dir, mask_dir, render_options
)
num_panos = len(pano_image_names)
max_workers = min(32, (os.cpu_count() or 2) - 1)
with tqdm(total=num_panos) as pbar:
with ThreadPoolExecutor(max_workers=max_workers) as thread_pool:
futures = [
thread_pool.submit(processor.process, pano_name)
for pano_name in pano_image_names
]
for future in as_completed(futures):
future.result()
pbar.update(1)
return processor.rig_config
def run(args):
pycolmap.set_random_seed(0)
# Define the paths.
image_dir = args.output_path / "images"
mask_dir = args.output_path / "masks"
image_dir.mkdir(exist_ok=True, parents=True)
mask_dir.mkdir(exist_ok=True, parents=True)
database_path = args.output_path / "database.db"
if database_path.exists():
database_path.unlink()
rec_path = args.output_path / "sparse"
rec_path.mkdir(exist_ok=True, parents=True)
# Search for input images.
pano_image_dir = args.input_image_path
pano_image_names = sorted(
p.relative_to(pano_image_dir).as_posix()
for p in pano_image_dir.rglob("*")
if not p.is_dir()
)
logging.info(f"Found {len(pano_image_names)} images in {pano_image_dir}.")
rig_config = render_perspective_images(
pano_image_names,
pano_image_dir,
image_dir,
mask_dir,
PANO_RENDER_OPTIONS[args.pano_render_type],
)
pycolmap.extract_features(
database_path,
image_dir,
reader_options={"mask_path": mask_dir},
camera_mode=pycolmap.CameraMode.PER_FOLDER,
)
with pycolmap.Database.open(database_path) as db:
pycolmap.apply_rig_config([rig_config], db)
matching_options = pycolmap.FeatureMatchingOptions()
# We have perfect sensor_from_rig poses (except for potential stitching
# artifacts by the spherical image provider), so we can perform geometric
# verification using rig constraints.
matching_options.rig_verification = True
# The images within a frame do not have overlap due to the provided masks.
matching_options.skip_image_pairs_in_same_frame = True
if args.matcher == "sequential":
pycolmap.match_sequential(
database_path,
pairing_options=pycolmap.SequentialPairingOptions(
loop_detection=True
),
matching_options=matching_options,
)
elif args.matcher == "exhaustive":
pycolmap.match_exhaustive(
database_path, matching_options=matching_options
)
elif args.matcher == "vocabtree":
pycolmap.match_vocabtree(
database_path, matching_options=matching_options
)
elif args.matcher == "spatial":
pycolmap.match_spatial(database_path, matching_options=matching_options)
else:
logging.fatal(f"Unknown matcher: {args.matcher}")
opts = pycolmap.IncrementalPipelineOptions(
ba_refine_sensor_from_rig=False,
ba_refine_focal_length=False,
ba_refine_principal_point=False,
ba_refine_extra_params=False,
)
recs = pycolmap.incremental_mapping(
database_path, image_dir, rec_path, opts
)
for idx, rec in recs.items():
logging.info(f"#{idx} {rec.summary()}")
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--input_image_path", type=Path, required=True)
parser.add_argument("--output_path", type=Path, required=True)
parser.add_argument(
"--matcher",
default="sequential",
choices=["sequential", "exhaustive", "vocabtree", "spatial"],
)
parser.add_argument(
"--pano_render_type",
default="overlapping",
choices=list(PANO_RENDER_OPTIONS.keys()),
)
run(parser.parse_args())
|