File size: 10,978 Bytes
f6228f9 |
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 |
# Ultralytics YOLO 🚀, AGPL-3.0 license
import itertools
from glob import glob
from math import ceil
from pathlib import Path
import cv2
import numpy as np
from PIL import Image
from tqdm import tqdm
from ultralytics.data.utils import exif_size, img2label_paths
from ultralytics.utils.checks import check_requirements
check_requirements("shapely")
from shapely.geometry import Polygon
def bbox_iof(polygon1, bbox2, eps=1e-6):
"""
Calculate Intersection over Foreground (IoF) between polygons and bounding boxes.
Args:
polygon1 (np.ndarray): Polygon coordinates, shape (n, 8).
bbox2 (np.ndarray): Bounding boxes, shape (n, 4).
eps (float, optional): Small value to prevent division by zero. Defaults to 1e-6.
Returns:
(np.ndarray): IoF scores, shape (n, 1) or (n, m) if bbox2 is (m, 4).
Note:
Polygon format: [x1, y1, x2, y2, x3, y3, x4, y4].
Bounding box format: [x_min, y_min, x_max, y_max].
"""
polygon1 = polygon1.reshape(-1, 4, 2)
lt_point = np.min(polygon1, axis=-2) # left-top
rb_point = np.max(polygon1, axis=-2) # right-bottom
bbox1 = np.concatenate([lt_point, rb_point], axis=-1)
lt = np.maximum(bbox1[:, None, :2], bbox2[..., :2])
rb = np.minimum(bbox1[:, None, 2:], bbox2[..., 2:])
wh = np.clip(rb - lt, 0, np.inf)
h_overlaps = wh[..., 0] * wh[..., 1]
left, top, right, bottom = (bbox2[..., i] for i in range(4))
polygon2 = np.stack([left, top, right, top, right, bottom, left, bottom], axis=-1).reshape(-1, 4, 2)
sg_polys1 = [Polygon(p) for p in polygon1]
sg_polys2 = [Polygon(p) for p in polygon2]
overlaps = np.zeros(h_overlaps.shape)
for p in zip(*np.nonzero(h_overlaps)):
overlaps[p] = sg_polys1[p[0]].intersection(sg_polys2[p[-1]]).area
unions = np.array([p.area for p in sg_polys1], dtype=np.float32)
unions = unions[..., None]
unions = np.clip(unions, eps, np.inf)
outputs = overlaps / unions
if outputs.ndim == 1:
outputs = outputs[..., None]
return outputs
def load_yolo_dota(data_root, split="train"):
"""
Load DOTA dataset.
Args:
data_root (str): Data root.
split (str): The split data set, could be train or val.
Notes:
The directory structure assumed for the DOTA dataset:
- data_root
- images
- train
- val
- labels
- train
- val
"""
assert split in {"train", "val"}, f"Split must be 'train' or 'val', not {split}."
im_dir = Path(data_root) / "images" / split
assert im_dir.exists(), f"Can't find {im_dir}, please check your data root."
im_files = glob(str(Path(data_root) / "images" / split / "*"))
lb_files = img2label_paths(im_files)
annos = []
for im_file, lb_file in zip(im_files, lb_files):
w, h = exif_size(Image.open(im_file))
with open(lb_file) as f:
lb = [x.split() for x in f.read().strip().splitlines() if len(x)]
lb = np.array(lb, dtype=np.float32)
annos.append(dict(ori_size=(h, w), label=lb, filepath=im_file))
return annos
def get_windows(im_size, crop_sizes=(1024,), gaps=(200,), im_rate_thr=0.6, eps=0.01):
"""
Get the coordinates of windows.
Args:
im_size (tuple): Original image size, (h, w).
crop_sizes (List(int)): Crop size of windows.
gaps (List(int)): Gap between crops.
im_rate_thr (float): Threshold of windows areas divided by image ares.
eps (float): Epsilon value for math operations.
"""
h, w = im_size
windows = []
for crop_size, gap in zip(crop_sizes, gaps):
assert crop_size > gap, f"invalid crop_size gap pair [{crop_size} {gap}]"
step = crop_size - gap
xn = 1 if w <= crop_size else ceil((w - crop_size) / step + 1)
xs = [step * i for i in range(xn)]
if len(xs) > 1 and xs[-1] + crop_size > w:
xs[-1] = w - crop_size
yn = 1 if h <= crop_size else ceil((h - crop_size) / step + 1)
ys = [step * i for i in range(yn)]
if len(ys) > 1 and ys[-1] + crop_size > h:
ys[-1] = h - crop_size
start = np.array(list(itertools.product(xs, ys)), dtype=np.int64)
stop = start + crop_size
windows.append(np.concatenate([start, stop], axis=1))
windows = np.concatenate(windows, axis=0)
im_in_wins = windows.copy()
im_in_wins[:, 0::2] = np.clip(im_in_wins[:, 0::2], 0, w)
im_in_wins[:, 1::2] = np.clip(im_in_wins[:, 1::2], 0, h)
im_areas = (im_in_wins[:, 2] - im_in_wins[:, 0]) * (im_in_wins[:, 3] - im_in_wins[:, 1])
win_areas = (windows[:, 2] - windows[:, 0]) * (windows[:, 3] - windows[:, 1])
im_rates = im_areas / win_areas
if not (im_rates > im_rate_thr).any():
max_rate = im_rates.max()
im_rates[abs(im_rates - max_rate) < eps] = 1
return windows[im_rates > im_rate_thr]
def get_window_obj(anno, windows, iof_thr=0.7):
"""Get objects for each window."""
h, w = anno["ori_size"]
label = anno["label"]
if len(label):
label[:, 1::2] *= w
label[:, 2::2] *= h
iofs = bbox_iof(label[:, 1:], windows)
# Unnormalized and misaligned coordinates
return [(label[iofs[:, i] >= iof_thr]) for i in range(len(windows))] # window_anns
else:
return [np.zeros((0, 9), dtype=np.float32) for _ in range(len(windows))] # window_anns
def crop_and_save(anno, windows, window_objs, im_dir, lb_dir, allow_background_images=True):
"""
Crop images and save new labels.
Args:
anno (dict): Annotation dict, including `filepath`, `label`, `ori_size` as its keys.
windows (list): A list of windows coordinates.
window_objs (list): A list of labels inside each window.
im_dir (str): The output directory path of images.
lb_dir (str): The output directory path of labels.
allow_background_images (bool): Whether to include background images without labels.
Notes:
The directory structure assumed for the DOTA dataset:
- data_root
- images
- train
- val
- labels
- train
- val
"""
im = cv2.imread(anno["filepath"])
name = Path(anno["filepath"]).stem
for i, window in enumerate(windows):
x_start, y_start, x_stop, y_stop = window.tolist()
new_name = f"{name}__{x_stop - x_start}__{x_start}___{y_start}"
patch_im = im[y_start:y_stop, x_start:x_stop]
ph, pw = patch_im.shape[:2]
label = window_objs[i]
if len(label) or allow_background_images:
cv2.imwrite(str(Path(im_dir) / f"{new_name}.jpg"), patch_im)
if len(label):
label[:, 1::2] -= x_start
label[:, 2::2] -= y_start
label[:, 1::2] /= pw
label[:, 2::2] /= ph
with open(Path(lb_dir) / f"{new_name}.txt", "w") as f:
for lb in label:
formatted_coords = [f"{coord:.6g}" for coord in lb[1:]]
f.write(f"{int(lb[0])} {' '.join(formatted_coords)}\n")
def split_images_and_labels(data_root, save_dir, split="train", crop_sizes=(1024,), gaps=(200,)):
"""
Split both images and labels.
Notes:
The directory structure assumed for the DOTA dataset:
- data_root
- images
- split
- labels
- split
and the output directory structure is:
- save_dir
- images
- split
- labels
- split
"""
im_dir = Path(save_dir) / "images" / split
im_dir.mkdir(parents=True, exist_ok=True)
lb_dir = Path(save_dir) / "labels" / split
lb_dir.mkdir(parents=True, exist_ok=True)
annos = load_yolo_dota(data_root, split=split)
for anno in tqdm(annos, total=len(annos), desc=split):
windows = get_windows(anno["ori_size"], crop_sizes, gaps)
window_objs = get_window_obj(anno, windows)
crop_and_save(anno, windows, window_objs, str(im_dir), str(lb_dir))
def split_trainval(data_root, save_dir, crop_size=1024, gap=200, rates=(1.0,)):
"""
Split train and val set of DOTA.
Notes:
The directory structure assumed for the DOTA dataset:
- data_root
- images
- train
- val
- labels
- train
- val
and the output directory structure is:
- save_dir
- images
- train
- val
- labels
- train
- val
"""
crop_sizes, gaps = [], []
for r in rates:
crop_sizes.append(int(crop_size / r))
gaps.append(int(gap / r))
for split in ["train", "val"]:
split_images_and_labels(data_root, save_dir, split, crop_sizes, gaps)
def split_test(data_root, save_dir, crop_size=1024, gap=200, rates=(1.0,)):
"""
Split test set of DOTA, labels are not included within this set.
Notes:
The directory structure assumed for the DOTA dataset:
- data_root
- images
- test
and the output directory structure is:
- save_dir
- images
- test
"""
crop_sizes, gaps = [], []
for r in rates:
crop_sizes.append(int(crop_size / r))
gaps.append(int(gap / r))
save_dir = Path(save_dir) / "images" / "test"
save_dir.mkdir(parents=True, exist_ok=True)
im_dir = Path(data_root) / "images" / "test"
assert im_dir.exists(), f"Can't find {im_dir}, please check your data root."
im_files = glob(str(im_dir / "*"))
for im_file in tqdm(im_files, total=len(im_files), desc="test"):
w, h = exif_size(Image.open(im_file))
windows = get_windows((h, w), crop_sizes=crop_sizes, gaps=gaps)
im = cv2.imread(im_file)
name = Path(im_file).stem
for window in windows:
x_start, y_start, x_stop, y_stop = window.tolist()
new_name = f"{name}__{x_stop - x_start}__{x_start}___{y_start}"
patch_im = im[y_start:y_stop, x_start:x_stop]
cv2.imwrite(str(save_dir / f"{new_name}.jpg"), patch_im)
if __name__ == "__main__":
split_trainval(data_root="DOTAv2", save_dir="DOTAv2-split")
split_test(data_root="DOTAv2", save_dir="DOTAv2-split")
|