Spaces:
Running
Running
File size: 10,932 Bytes
499e141 |
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 |
import argparse
import torch
from pathlib import Path
from typing import Dict, List, Union, Optional
import h5py
from types import SimpleNamespace
import cv2
import numpy as np
from tqdm import tqdm
import pprint
import collections.abc as collections
import PIL.Image
import glob
from . import extractors, logger
from .utils.base_model import dynamic_load
from .utils.parsers import parse_image_lists
from .utils.io import read_image, list_h5_names
'''
A set of standard configurations that can be directly selected from the command
line using their name. Each is a dictionary with the following entries:
- output: the name of the feature file that will be generated.
- model: the model configuration, as passed to a feature extractor.
- preprocessing: how to preprocess the images read from disk.
'''
confs = {
'gim_superpoint': {
'output': 'feats-gim-superpoint-n2048-r1920',
'model': {
'name': 'superpoint',
'nms_radius': 3,
'max_keypoints': 2048,
},
'preprocessing': {
'grayscale': True,
'resize_max': 1920,
},
},
'superpoint_aachen': {
'output': 'feats-superpoint-n4096-r1024',
'model': {
'name': 'superpoint',
'nms_radius': 3,
'max_keypoints': 4096,
},
'preprocessing': {
'grayscale': True,
'resize_max': 1024,
},
},
# Resize images to 1600px even if they are originally smaller.
# Improves the keypoint localization if the images are of good quality.
'superpoint_max': {
'output': 'feats-superpoint-n4096-rmax1600',
'model': {
'name': 'superpoint',
'nms_radius': 3,
'max_keypoints': 4096,
},
'preprocessing': {
'grayscale': True,
'resize_max': 1600,
'resize_force': True,
},
},
'superpoint_inloc': {
'output': 'feats-superpoint-n4096-r1600',
'model': {
'name': 'superpoint',
'nms_radius': 4,
'max_keypoints': 4096,
},
'preprocessing': {
'grayscale': True,
'resize_max': 2048,
},
},
'r2d2': {
'output': 'feats-r2d2-n5000-r1024',
'model': {
'name': 'r2d2',
'max_keypoints': 5000,
},
'preprocessing': {
'grayscale': False,
'resize_max': 1024,
},
},
'd2net-ss': {
'output': 'feats-d2net-ss',
'model': {
'name': 'd2net',
'multiscale': False,
},
'preprocessing': {
'grayscale': False,
'resize_max': 1600,
},
},
'sift': {
'output': 'feats-sift',
'model': {
'name': 'dog'
},
'preprocessing': {
'grayscale': True,
'resize_max': 1600,
},
},
'sosnet': {
'output': 'feats-sosnet',
'model': {
'name': 'dog',
'descriptor': 'sosnet'
},
'preprocessing': {
'grayscale': True,
'resize_max': 1600,
},
},
'disk': {
'output': 'feats-disk',
'model': {
'name': 'disk',
'max_keypoints': 5000,
},
'preprocessing': {
'grayscale': False,
'resize_max': 1600,
},
},
# Global descriptors
'dir': {
'output': 'global-feats-dir',
'model': {'name': 'dir'},
'preprocessing': {'resize_max': 1024},
},
'netvlad': {
'output': 'global-feats-netvlad',
'model': {'name': 'netvlad'},
'preprocessing': {'resize_max': 1024},
},
'openibl': {
'output': 'global-feats-openibl',
'model': {'name': 'openibl'},
'preprocessing': {'resize_max': 1024},
},
'cosplace': {
'output': 'global-feats-cosplace',
'model': {'name': 'cosplace'},
'preprocessing': {'resize_max': 1024},
}
}
def resize_image(image, size, interp):
if interp.startswith('cv2_'):
interp = getattr(cv2, 'INTER_'+interp[len('cv2_'):].upper())
h, w = image.shape[:2]
if interp == cv2.INTER_AREA and (w < size[0] or h < size[1]):
interp = cv2.INTER_LINEAR
resized = cv2.resize(image, size, interpolation=interp)
elif interp.startswith('pil_'):
interp = getattr(PIL.Image, interp[len('pil_'):].upper())
resized = PIL.Image.fromarray(image.astype(np.uint8))
resized = resized.resize(size, resample=interp)
resized = np.asarray(resized, dtype=image.dtype)
else:
raise ValueError(
f'Unknown interpolation {interp}.')
return resized
class ImageDataset(torch.utils.data.Dataset):
default_conf = {
'globs': ['*.jpg', '*.png', '*.jpeg', '*.JPG', '*.PNG'],
'grayscale': False,
'resize_max': None,
'resize_force': False,
'interpolation': 'cv2_area', # pil_linear is more accurate but slower
}
def __init__(self, root, conf, paths=None):
self.conf = conf = SimpleNamespace(**{**self.default_conf, **conf})
self.root = root
if paths is None:
paths = []
for g in conf.globs:
paths += glob.glob(
(Path(root) / '**' / g).as_posix(), recursive=True)
if len(paths) == 0:
raise ValueError(f'Could not find any image in root: {root}.')
paths = sorted(set(paths))
self.names = [Path(p).relative_to(root).as_posix() for p in paths]
logger.info(f'Found {len(self.names)} images in root {root}.')
else:
if isinstance(paths, (Path, str)):
self.names = parse_image_lists(paths)
elif isinstance(paths, collections.Iterable):
self.names = [p.as_posix() if isinstance(p, Path) else p
for p in paths]
else:
raise ValueError(f'Unknown format for path argument {paths}.')
for name in self.names:
if not (root / name).exists():
raise ValueError(
f'Image {name} does not exists in root: {root}.')
def __getitem__(self, idx):
name = self.names[idx]
image = read_image(self.root / name, self.conf.grayscale)
image = image.astype(np.float32)
size = image.shape[:2][::-1]
if self.conf.resize_max and (self.conf.resize_force
or max(size) > self.conf.resize_max):
scale = self.conf.resize_max / max(size)
size_new = tuple(int(round(x*scale)) for x in size)
image = resize_image(image, size_new, self.conf.interpolation)
if self.conf.grayscale:
image = image[None]
else:
image = image.transpose((2, 0, 1)) # HxWxC to CxHxW
image = image / 255.
data = {
'image': image,
'original_size': np.array(size),
}
return data
def __len__(self):
return len(self.names)
@torch.no_grad()
def main(conf: Dict,
image_dir: Path,
export_dir: Optional[Path] = None,
as_half: bool = True,
image_list: Optional[Union[Path, List[str]]] = None,
feature_path: Optional[Path] = None,
overwrite: bool = False,
model=None) -> Path:
logger.info('Extracting local features with configuration:'
f'\n{pprint.pformat(conf)}')
dataset = ImageDataset(image_dir, conf['preprocessing'], image_list)
if feature_path is None:
feature_path = Path(export_dir, conf['output']+'.h5')
feature_path.parent.mkdir(exist_ok=True, parents=True)
skip_names = set(list_h5_names(feature_path)
if feature_path.exists() and not overwrite else ())
dataset.names = [n for n in dataset.names if n not in skip_names]
if len(dataset.names) == 0:
logger.info('Skipping the extraction.')
return feature_path
device = 'cuda' if torch.cuda.is_available() else 'cpu'
if model is None:
Model = dynamic_load(extractors, conf['model']['name'])
model = Model(conf['model'])
model = model.eval().to(device)
loader = torch.utils.data.DataLoader(
dataset, num_workers=1, shuffle=False, pin_memory=True)
for idx, data in enumerate(tqdm(loader)):
name = dataset.names[idx]
pred = model({'image': data['image'].to(device, non_blocking=True)})
pred = {k: v[0].cpu().numpy() for k, v in pred.items()}
pred['image_size'] = original_size = data['original_size'][0].numpy()
if 'keypoints' in pred:
size = np.array(data['image'].shape[-2:][::-1])
scales = (original_size / size).astype(np.float32)
pred['keypoints'] = (pred['keypoints'] + .5) * scales[None] - .5
if 'scales' in pred:
pred['scales'] *= scales.mean()
# add keypoint uncertainties scaled to the original resolution
uncertainty = getattr(model, 'detection_noise', 1) * scales.mean()
if as_half:
for k in pred:
dt = pred[k].dtype
if (dt == np.float32) and (dt != np.float16):
pred[k] = pred[k].astype(np.float16)
with h5py.File(str(feature_path), 'a', libver='latest') as fd:
try:
if name in fd:
del fd[name]
grp = fd.create_group(name)
for k, v in pred.items():
grp.create_dataset(k, data=v)
if 'keypoints' in pred:
grp['keypoints'].attrs['uncertainty'] = uncertainty
except OSError as error:
if 'No space left on device' in error.args[0]:
logger.error(
'Out of disk space: storing features on disk can take '
'significant space, did you enable the as_half flag?')
del grp, fd[name]
raise error
del pred
logger.info('Finished exporting features.')
return feature_path
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--image_dir', type=Path, required=True)
parser.add_argument('--export_dir', type=Path, required=True)
parser.add_argument('--conf', type=str, default='superpoint_aachen',
choices=list(confs.keys()))
parser.add_argument('--as_half', action='store_true')
parser.add_argument('--image_list', type=Path)
parser.add_argument('--feature_path', type=Path)
args = parser.parse_args()
main(confs[args.conf], args.image_dir, args.export_dir, args.as_half)
|