File size: 7,618 Bytes
550b1d5 | 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 | """
WiderFace Dataset Loader.
WIDER FACE (Yang et al., 2016):
- 32,203 images, 393,703 annotated face bounding boxes
- Split: 40% train (12,880), 10% val (3,226), 50% test (labels not public)
- 3 difficulty levels: Easy, Medium, Hard
- Annotations include: bbox, blur, expression, illumination, occlusion, pose, invalid
Directory structure expected:
wider_face/
├── WIDER_train/
│ └── images/
│ ├── 0--Parade/
│ ├── 1--Handshaking/
│ └── ...
├── WIDER_val/
│ └── images/
│ └── ...
├── wider_face_split/
│ ├── wider_face_train_bbx_gt.txt
│ ├── wider_face_val_bbx_gt.txt
│ └── ...
└── retinaface_gt/ (optional, for landmarks)
├── train/
│ └── label.txt
└── val/
└── label.txt
"""
import os
import numpy as np
import cv2
from typing import List, Dict, Optional, Tuple, Callable
import torch
from torch.utils.data import Dataset
class WiderFaceDataset(Dataset):
"""
WIDER FACE dataset with support for:
- Standard WiderFace bbox annotations
- RetinaFace-format 5-point landmark annotations
- Filtering invalid/tiny faces
- On-the-fly augmentation
"""
def __init__(self,
root_dir: str,
split: str = 'train',
transform: Optional[Callable] = None,
min_face_size: int = 2,
use_landmarks: bool = False,
annotation_format: str = 'widerface'):
"""
Args:
root_dir: Path to wider_face/ directory
split: 'train' or 'val'
transform: Augmentation callable
min_face_size: Minimum face size to keep (pixels)
use_landmarks: Load 5-point landmarks (requires retinaface_gt/)
annotation_format: 'widerface' (standard) or 'retinaface' (with landmarks)
"""
self.root_dir = root_dir
self.split = split
self.transform = transform
self.min_face_size = min_face_size
self.use_landmarks = use_landmarks
if annotation_format == 'retinaface' and use_landmarks:
self.samples = self._load_retinaface_annotations()
else:
self.samples = self._load_widerface_annotations()
print(f"[WiderFace {split}] Loaded {len(self.samples)} images")
def _load_widerface_annotations(self) -> List[Dict]:
"""Load standard WiderFace bbox annotations."""
ann_file = os.path.join(
self.root_dir, 'wider_face_split',
f'wider_face_{self.split}_bbx_gt.txt'
)
img_dir = os.path.join(self.root_dir, f'WIDER_{self.split}', 'images')
samples = []
with open(ann_file, 'r') as f:
while True:
filename = f.readline().strip()
if not filename:
break
num_faces = int(f.readline().strip())
boxes = []
for _ in range(max(num_faces, 1)):
line = f.readline().strip()
parts = list(map(float, line.split()))
if num_faces == 0:
continue # Skip placeholder line for 0-face images
x, y, w, h = parts[0], parts[1], parts[2], parts[3]
# Filter tiny/invalid faces
if w < self.min_face_size or h < self.min_face_size:
continue
# Convert to x1, y1, x2, y2
boxes.append([x, y, x + w, y + h])
if boxes:
samples.append({
'image_path': os.path.join(img_dir, filename),
'boxes': np.array(boxes, dtype=np.float32),
'filename': filename,
})
return samples
def _load_retinaface_annotations(self) -> List[Dict]:
"""Load RetinaFace-format annotations with 5-point landmarks."""
ann_file = os.path.join(
self.root_dir, 'retinaface_gt', self.split, 'label.txt'
)
img_dir = os.path.join(self.root_dir, f'WIDER_{self.split}', 'images')
samples = []
current_file = None
current_boxes = []
current_lmks = []
with open(ann_file, 'r') as f:
for line in f:
line = line.strip()
if line.startswith('#'):
# Save previous image
if current_file and current_boxes:
samples.append({
'image_path': os.path.join(img_dir, current_file),
'boxes': np.array(current_boxes, dtype=np.float32),
'landmarks': np.array(current_lmks, dtype=np.float32),
'filename': current_file,
})
current_file = line[2:].strip()
current_boxes = []
current_lmks = []
else:
parts = list(map(float, line.split()))
if len(parts) >= 4:
x, y, w, h = parts[0], parts[1], parts[2], parts[3]
if w < self.min_face_size or h < self.min_face_size:
continue
current_boxes.append([x, y, x + w, y + h])
if len(parts) >= 14:
# 5 landmarks: (x1,y1, x2,y2, x3,y3, x4,y4, x5,y5)
lmk = parts[4:14]
current_lmks.append(lmk)
else:
current_lmks.append([-1]*10) # Invalid landmarks
# Save last image
if current_file and current_boxes:
samples.append({
'image_path': os.path.join(img_dir, current_file),
'boxes': np.array(current_boxes, dtype=np.float32),
'landmarks': np.array(current_lmks, dtype=np.float32),
'filename': current_file,
})
return samples
def __len__(self) -> int:
return len(self.samples)
def __getitem__(self, idx: int) -> Dict:
sample = self.samples[idx]
# Load image
img = cv2.imread(sample['image_path'])
if img is None:
raise IOError(f"Failed to load image: {sample['image_path']}")
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
boxes = sample['boxes'].copy()
landmarks = sample.get('landmarks', np.zeros((boxes.shape[0], 10), dtype=np.float32)).copy()
# Apply augmentation
if self.transform:
result = self.transform(img, boxes, landmarks)
img, boxes, landmarks = result['image'], result['boxes'], result['landmarks']
# Convert to tensors
img_tensor = torch.from_numpy(img.transpose(2, 0, 1)).float()
boxes_tensor = torch.from_numpy(boxes).float()
target = {
'boxes': boxes_tensor,
'labels': torch.ones(boxes_tensor.shape[0], dtype=torch.long),
}
if self.use_landmarks:
target['landmarks'] = torch.from_numpy(landmarks).float()
return img_tensor, target
@staticmethod
def collate_fn(batch):
"""Custom collate for variable-length targets."""
images = torch.stack([item[0] for item in batch])
targets = [item[1] for item in batch]
return images, targets
|