File size: 11,410 Bytes
99f86f1 | 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 | """
DAWN Dataset Preparation Script
Downloads from HuggingFace Hub, converts to YOLO format, applies augmentation
for minority classes, and creates train/val/test splits.
"""
import os
import json
import random
import shutil
import numpy as np
from pathlib import Path
from datasets import load_dataset
from PIL import Image, ImageOps
# βββ Configuration βββββββββββββββββββββββββββββββββββββββββββββββββββ
DATASET_ROOT = "/app/dawn_dataset"
SEED = 42
TRAIN_RATIO = 0.60
VAL_RATIO = 0.20
TEST_RATIO = 0.20
# Class mapping matching user's specification
CLASS_NAMES = ['Bicycle', 'Bus', 'Car', 'Motorcycle', 'Pedestrian', 'Truck']
# Map from dataset class_name to our index
CLASS_MAP = {
'Bicycle': 0,
'Bus': 1,
'Car': 2,
'Motorcycle': 3,
'Pedestrian': 4, 'Person': 4, 'Cyclist': 4,
'Truck': 5,
}
random.seed(SEED)
np.random.seed(SEED)
def setup_dirs():
"""Create YOLO directory structure."""
for split in ['train', 'val', 'test']:
os.makedirs(f"{DATASET_ROOT}/images/{split}", exist_ok=True)
os.makedirs(f"{DATASET_ROOT}/labels/{split}", exist_ok=True)
def convert_to_yolo(objects, img_w, img_h):
"""Convert absolute bbox annotations to YOLO normalized format."""
labels = []
for obj in objects:
cls_name = obj['class_name']
if cls_name not in CLASS_MAP:
print(f" WARNING: Unknown class '{cls_name}', skipping")
continue
cls_id = CLASS_MAP[cls_name]
x_min = obj['x_min']
y_min = obj['y_min']
w = obj['width']
h = obj['height']
# Convert to YOLO format: cx, cy, w, h (normalized)
cx = (x_min + w / 2) / img_w
cy = (y_min + h / 2) / img_h
nw = w / img_w
nh = h / img_h
# Clip to [0, 1]
cx = max(0, min(1, cx))
cy = max(0, min(1, cy))
nw = max(0, min(1, nw))
nh = max(0, min(1, nh))
if nw > 0.001 and nh > 0.001: # skip degenerate boxes
labels.append(f"{cls_id} {cx:.6f} {cy:.6f} {nw:.6f} {nh:.6f}")
return labels
def save_image_and_label(image, labels, img_name, split):
"""Save image and YOLO label file."""
img_path = f"{DATASET_ROOT}/images/{split}/{img_name}.jpg"
lbl_path = f"{DATASET_ROOT}/labels/{split}/{img_name}.txt"
if isinstance(image, Image.Image):
image.save(img_path, quality=95)
else:
image.save(img_path)
with open(lbl_path, 'w') as f:
f.write('\n'.join(labels))
def augment_mirror(image, labels_raw, img_w, img_h):
"""Horizontal flip augmentation with bbox adjustment."""
flipped = ImageOps.mirror(image)
new_labels = []
for lbl in labels_raw:
parts = lbl.split()
cls_id = parts[0]
cx, cy, w, h = float(parts[1]), float(parts[2]), float(parts[3]), float(parts[4])
new_cx = 1.0 - cx
new_labels.append(f"{cls_id} {new_cx:.6f} {cy:.6f} {w:.6f} {h:.6f}")
return flipped, new_labels
def augment_rotate(image, labels_raw, img_w, img_h, angle=90):
"""Rotation augmentation (90, 180, 270 degrees) with bbox adjustment."""
if angle == 90:
rotated = image.transpose(Image.ROTATE_90)
new_labels = []
for lbl in labels_raw:
parts = lbl.split()
cls_id = parts[0]
cx, cy, w, h = float(parts[1]), float(parts[2]), float(parts[3]), float(parts[4])
new_cx, new_cy = cy, 1.0 - cx
new_w, new_h = h, w
new_labels.append(f"{cls_id} {new_cx:.6f} {new_cy:.6f} {new_w:.6f} {new_h:.6f}")
elif angle == 180:
rotated = image.transpose(Image.ROTATE_180)
new_labels = []
for lbl in labels_raw:
parts = lbl.split()
cls_id = parts[0]
cx, cy, w, h = float(parts[1]), float(parts[2]), float(parts[3]), float(parts[4])
new_cx, new_cy = 1.0 - cx, 1.0 - cy
new_labels.append(f"{cls_id} {new_cx:.6f} {new_cy:.6f} {w:.6f} {h:.6f}")
elif angle == 270:
rotated = image.transpose(Image.ROTATE_270)
new_labels = []
for lbl in labels_raw:
parts = lbl.split()
cls_id = parts[0]
cx, cy, w, h = float(parts[1]), float(parts[2]), float(parts[3]), float(parts[4])
new_cx, new_cy = 1.0 - cy, cx
new_w, new_h = h, w
new_labels.append(f"{cls_id} {new_cx:.6f} {new_cy:.6f} {new_w:.6f} {new_h:.6f}")
else:
return image, labels_raw
return rotated, new_labels
def main():
print("=" * 60)
print("DAWN Dataset Preparation Pipeline")
print("=" * 60)
setup_dirs()
# Load dataset from HF Hub
print("\n[1/5] Loading DAWN dataset from HuggingFace Hub...")
ds = load_dataset("Maxim37/dawn-dataset")
print(f" Train split: {len(ds['train'])} images")
print(f" Val split: {len(ds['val'])} images")
# Combine all data for re-splitting
all_samples = []
for split_name in ['train', 'val']:
for idx, sample in enumerate(ds[split_name]):
all_samples.append(sample)
print(f" Total images: {len(all_samples)}")
# βββ Phase 1: Convert all images to YOLO format ββββββββββββββββββ
print("\n[2/5] Converting annotations to YOLO format...")
converted = []
class_counts = {name: 0 for name in CLASS_NAMES}
for i, sample in enumerate(all_samples):
img = sample['image']
if not isinstance(img, Image.Image):
continue
img_w = sample['width']
img_h = sample['height']
image_id = sample['image_id']
objects = sample['objects']
labels = convert_to_yolo(objects, img_w, img_h)
if len(labels) == 0:
continue
# Count classes in this image
img_classes = set()
for lbl in labels:
cls_id = int(lbl.split()[0])
class_counts[CLASS_NAMES[cls_id]] += 1
img_classes.add(cls_id)
converted.append({
'image': img,
'labels': labels,
'image_id': image_id,
'img_classes': img_classes,
'img_w': img_w,
'img_h': img_h,
})
if (i + 1) % 100 == 0:
print(f" Processed {i + 1}/{len(all_samples)} images...")
print(f" Successfully converted: {len(converted)} images")
print(f"\n Class distribution (before augmentation):")
for name, count in class_counts.items():
print(f" {name}: {count} instances")
# βββ Phase 2: Identify minority classes & augment βββββββββββββββββ
print("\n[3/5] Augmenting minority classes...")
total_instances = sum(class_counts.values())
mean_count = total_instances / len(CLASS_NAMES)
# Classes below mean are minority
minority_classes = set()
for name, count in class_counts.items():
if count < mean_count * 0.5: # Less than 50% of mean
minority_classes.add(CLASS_NAMES.index(name))
print(f" Minority class: {name} ({count} instances)")
# Augment images containing minority classes
augmented_samples = []
for sample in converted:
has_minority = bool(sample['img_classes'] & minority_classes)
if has_minority:
img = sample['image']
labels = sample['labels']
img_w = sample['img_w']
img_h = sample['img_h']
base_id = sample['image_id']
# Mirror augmentation
mir_img, mir_labels = augment_mirror(img, labels, img_w, img_h)
augmented_samples.append({
'image': mir_img,
'labels': mir_labels,
'image_id': f"{base_id}_mirror",
'img_classes': sample['img_classes'],
})
# Rotation augmentations (90Β° and 180Β°)
for angle in [90, 180]:
rot_img, rot_labels = augment_rotate(img, labels, img_w, img_h, angle)
augmented_samples.append({
'image': rot_img,
'labels': rot_labels,
'image_id': f"{base_id}_rot{angle}",
'img_classes': sample['img_classes'],
})
all_data = converted + augmented_samples
print(f" Original images: {len(converted)}")
print(f" Augmented images: {len(augmented_samples)}")
print(f" Total images: {len(all_data)}")
# βββ Phase 3: Split into train/val/test βββββββββββββββββββββββββββ
print("\n[4/5] Splitting into train/val/test (60/20/20)...")
random.shuffle(all_data)
n = len(all_data)
n_train = int(n * TRAIN_RATIO)
n_val = int(n * VAL_RATIO)
splits = {
'train': all_data[:n_train],
'val': all_data[n_train:n_train + n_val],
'test': all_data[n_train + n_val:],
}
for split_name, split_data in splits.items():
print(f" {split_name}: {len(split_data)} images")
# βββ Phase 4: Save everything βββββββββββββββββββββββββββββββββββββ
print("\n[5/5] Saving images and labels...")
split_class_counts = {s: {n: 0 for n in CLASS_NAMES} for s in ['train', 'val', 'test']}
for split_name, split_data in splits.items():
for i, sample in enumerate(split_data):
img_name = f"{split_name}_{i:05d}"
save_image_and_label(sample['image'], sample['labels'], img_name, split_name)
for lbl in sample['labels']:
cls_id = int(lbl.split()[0])
split_class_counts[split_name][CLASS_NAMES[cls_id]] += 1
if (i + 1) % 200 == 0:
print(f" [{split_name}] Saved {i + 1}/{len(split_data)}")
# Print final statistics
print("\n" + "=" * 60)
print("FINAL DATASET STATISTICS")
print("=" * 60)
for split_name in ['train', 'val', 'test']:
print(f"\n {split_name.upper()}:")
for cls_name, count in split_class_counts[split_name].items():
print(f" {cls_name}: {count} instances")
# βββ Create dataset YAML βββββββββββββββββββββββββββββββββββββββββ
yaml_content = f"""# DAWN Dataset - Vehicle Detection in Adverse Weather
path: {DATASET_ROOT}
train: images/train
val: images/val
test: images/test
nc: {len(CLASS_NAMES)}
names: {CLASS_NAMES}
"""
yaml_path = f"{DATASET_ROOT}/dataset.yaml"
with open(yaml_path, 'w') as f:
f.write(yaml_content)
print(f"\n Dataset YAML saved to: {yaml_path}")
# Save metadata
metadata = {
'total_images': len(all_data),
'original_images': len(converted),
'augmented_images': len(augmented_samples),
'splits': {s: len(d) for s, d in splits.items()},
'class_names': CLASS_NAMES,
'class_counts': {s: split_class_counts[s] for s in ['train', 'val', 'test']},
}
with open(f"{DATASET_ROOT}/metadata.json", 'w') as f:
json.dump(metadata, f, indent=2)
print("\nβ
Dataset preparation complete!")
print(f" Root: {DATASET_ROOT}")
return metadata
if __name__ == "__main__":
main()
|