Spaces:
Sleeping
Sleeping
File size: 10,886 Bytes
8b4d6a8 | 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 | """
Annotation corruption strategies for the Annotation QA Environment.
Takes gold-standard annotations and systematically corrupts them to create
training data with known errors. The corruption is deterministic given a seed.
Corruption types by difficulty:
- Task 1 (Easy): Obvious bbox errors — expand, shift, delete, add spurious
- Task 2 (Medium): bbox + class errors — similar class confusion, boundary errors
- Task 3 (Hard): Cross-image inconsistencies + subtle errors
"""
import copy
import random
from typing import Dict, List, Tuple
# Class confusion maps — used for "similar class" corruption
SIMILAR_CLASSES: Dict[str, List[str]] = {
"car": ["truck", "van"],
"truck": ["car", "van"],
"van": ["car", "truck"],
"person": ["cyclist"],
"cyclist": ["person"],
"dog": ["cat"],
"cat": ["dog"],
"bicycle": ["motorcycle"],
"motorcycle": ["bicycle"],
"tree": ["bush"],
"bush": ["tree"],
"building": ["house"],
"house": ["building"],
"traffic_light": ["street_light"],
"street_light": ["traffic_light"],
"bench": ["chair"],
"chair": ["bench"],
}
# Completely different classes for "wrong category" corruption
ALL_CLASSES = [
"car", "truck", "person", "bicycle", "dog", "cat",
"tree", "building", "traffic_light", "bench",
]
def _clamp(val: float, lo: float = 0.0, hi: float = 1.0) -> float:
return max(lo, min(hi, val))
def _clamp_bbox(bbox: List[float]) -> List[float]:
"""Ensure bbox stays within [0, 1] image bounds."""
x, y, w, h = bbox
x = _clamp(x)
y = _clamp(y)
w = _clamp(w, 0.02, 1.0 - x)
h = _clamp(h, 0.02, 1.0 - y)
return [round(x, 4), round(y, 4), round(w, 4), round(h, 4)]
def expand_bbox(bbox: List[float], factor: float) -> List[float]:
"""Expand a bbox by a factor (e.g., 1.5 = 50% larger)."""
x, y, w, h = bbox
cx, cy = x + w / 2, y + h / 2
new_w, new_h = w * factor, h * factor
new_x = cx - new_w / 2
new_y = cy - new_h / 2
return _clamp_bbox([new_x, new_y, new_w, new_h])
def shift_bbox(bbox: List[float], dx_frac: float, dy_frac: float) -> List[float]:
"""Shift a bbox by a fraction of its size."""
x, y, w, h = bbox
new_x = x + w * dx_frac
new_y = y + h * dy_frac
return _clamp_bbox([new_x, new_y, w, h])
def shrink_bbox(bbox: List[float], factor: float) -> List[float]:
"""Shrink a bbox (factor < 1.0)."""
return expand_bbox(bbox, factor)
def generate_spurious_annotation(
existing_bboxes: List[List[float]], rng: random.Random
) -> Dict:
"""Generate a random annotation that doesn't overlap much with existing ones."""
for _ in range(20): # try up to 20 times
w = rng.uniform(0.05, 0.20)
h = rng.uniform(0.05, 0.20)
x = rng.uniform(0.0, 1.0 - w)
y = rng.uniform(0.0, 1.0 - h)
bbox = [round(x, 4), round(y, 4), round(w, 4), round(h, 4)]
# Check it doesn't overlap too much with existing
from .grader import compute_iou
max_iou = max(
(compute_iou(bbox, eb) for eb in existing_bboxes), default=0.0
)
if max_iou < 0.3:
cls = rng.choice(ALL_CLASSES)
return {"bbox": bbox, "class_label": cls}
# Fallback: place it anyway
return {
"bbox": [round(rng.uniform(0.0, 0.8), 4), round(rng.uniform(0.0, 0.8), 4), 0.1, 0.1],
"class_label": rng.choice(ALL_CLASSES),
}
def corrupt_annotations(
gold_annotations: List[Dict],
difficulty: str,
seed: int,
) -> Tuple[List[Dict], List[str]]:
"""
Corrupt gold annotations based on difficulty level.
Returns:
(corrupted_annotations, corruption_log)
corruption_log: list of strings describing what was corrupted (for debugging)
"""
rng = random.Random(seed)
corrupted = copy.deepcopy(gold_annotations)
log = []
if difficulty == "easy":
# Task 1: Obvious bbox errors only (no class changes)
corruption_rate = 0.35
n_corrupt = max(1, int(len(corrupted) * corruption_rate))
indices = list(range(len(corrupted)))
rng.shuffle(indices)
corrupt_indices = indices[:n_corrupt]
for idx in corrupt_indices:
action = rng.choice(["expand", "shift", "shrink", "delete"])
ann = corrupted[idx]
if action == "expand":
factor = rng.uniform(1.5, 2.5)
ann["bbox"] = expand_bbox(ann["bbox"], factor)
log.append(f"Expanded ann {ann['id']} by {factor:.1f}x")
elif action == "shift":
dx = rng.uniform(-0.4, 0.4)
dy = rng.uniform(-0.4, 0.4)
ann["bbox"] = shift_bbox(ann["bbox"], dx, dy)
log.append(f"Shifted ann {ann['id']} by ({dx:.2f}, {dy:.2f})")
elif action == "shrink":
factor = rng.uniform(0.3, 0.6)
ann["bbox"] = shrink_bbox(ann["bbox"], factor)
log.append(f"Shrunk ann {ann['id']} by {factor:.1f}x")
elif action == "delete":
log.append(f"Deleted ann {ann['id']} ({ann['class_label']})")
corrupted[idx] = None # mark for removal
# Remove deleted
corrupted = [a for a in corrupted if a is not None]
# Add 2-3 spurious annotations
existing_bboxes = [a["bbox"] for a in corrupted]
n_spurious = rng.randint(2, 3)
next_id = max((a["id"] for a in corrupted), default=0) + 1
for i in range(n_spurious):
spur = generate_spurious_annotation(existing_bboxes, rng)
spur["id"] = next_id + i
corrupted.append(spur)
existing_bboxes.append(spur["bbox"])
log.append(f"Added spurious ann {spur['id']} ({spur['class_label']})")
elif difficulty == "medium":
# Task 2: bbox errors + class confusion
corruption_rate = 0.30
n_corrupt = max(2, int(len(corrupted) * corruption_rate))
indices = list(range(len(corrupted)))
rng.shuffle(indices)
corrupt_indices = indices[:n_corrupt]
for idx in corrupt_indices:
action = rng.choice([
"expand", "shift", "wrong_similar_class",
"wrong_different_class", "delete",
])
ann = corrupted[idx]
if action == "expand":
factor = rng.uniform(1.3, 2.0)
ann["bbox"] = expand_bbox(ann["bbox"], factor)
log.append(f"Expanded ann {ann['id']} by {factor:.1f}x")
elif action == "shift":
dx = rng.uniform(-0.3, 0.3)
dy = rng.uniform(-0.3, 0.3)
ann["bbox"] = shift_bbox(ann["bbox"], dx, dy)
log.append(f"Shifted ann {ann['id']}")
elif action == "wrong_similar_class":
old_cls = ann["class_label"]
similar = SIMILAR_CLASSES.get(old_cls, [])
if similar:
new_cls = rng.choice(similar)
ann["class_label"] = new_cls
log.append(f"Changed ann {ann['id']} class: {old_cls} → {new_cls}")
else:
# Fallback to a different class
candidates = [c for c in ALL_CLASSES if c != old_cls]
ann["class_label"] = rng.choice(candidates)
log.append(f"Changed ann {ann['id']} class: {old_cls} → {ann['class_label']}")
elif action == "wrong_different_class":
old_cls = ann["class_label"]
candidates = [c for c in ALL_CLASSES if c != old_cls]
ann["class_label"] = rng.choice(candidates)
log.append(f"Changed ann {ann['id']} class: {old_cls} → {ann['class_label']} (wrong category)")
elif action == "delete":
log.append(f"Deleted ann {ann['id']} ({ann['class_label']})")
corrupted[idx] = None
corrupted = [a for a in corrupted if a is not None]
# Add 3-4 spurious
existing_bboxes = [a["bbox"] for a in corrupted]
n_spurious = rng.randint(3, 4)
next_id = max((a["id"] for a in corrupted), default=0) + 1
for i in range(n_spurious):
spur = generate_spurious_annotation(existing_bboxes, rng)
spur["id"] = next_id + i
corrupted.append(spur)
existing_bboxes.append(spur["bbox"])
log.append(f"Added spurious ann {spur['id']} ({spur['class_label']})")
elif difficulty == "hard":
# Task 3: Subtle errors + class confusion + some bbox
corruption_rate = 0.25
n_corrupt = max(2, int(len(corrupted) * corruption_rate))
indices = list(range(len(corrupted)))
rng.shuffle(indices)
corrupt_indices = indices[:n_corrupt]
for idx in corrupt_indices:
action = rng.choice([
"subtle_shift", "wrong_similar_class",
"wrong_similar_class", "delete", "subtle_expand",
])
ann = corrupted[idx]
if action == "subtle_shift":
dx = rng.uniform(-0.15, 0.15)
dy = rng.uniform(-0.15, 0.15)
ann["bbox"] = shift_bbox(ann["bbox"], dx, dy)
log.append(f"Subtly shifted ann {ann['id']}")
elif action == "subtle_expand":
factor = rng.uniform(1.15, 1.4)
ann["bbox"] = expand_bbox(ann["bbox"], factor)
log.append(f"Subtly expanded ann {ann['id']}")
elif action == "wrong_similar_class":
old_cls = ann["class_label"]
similar = SIMILAR_CLASSES.get(old_cls, [])
if similar:
new_cls = rng.choice(similar)
ann["class_label"] = new_cls
log.append(f"Changed ann {ann['id']}: {old_cls} → {new_cls} (similar)")
else:
candidates = [c for c in ALL_CLASSES if c != old_cls]
ann["class_label"] = rng.choice(candidates)
log.append(f"Changed ann {ann['id']}: {old_cls} → {ann['class_label']}")
elif action == "delete":
log.append(f"Deleted ann {ann['id']}")
corrupted[idx] = None
corrupted = [a for a in corrupted if a is not None]
# Add 2-3 spurious
existing_bboxes = [a["bbox"] for a in corrupted]
n_spurious = rng.randint(2, 3)
next_id = max((a["id"] for a in corrupted), default=0) + 1
for i in range(n_spurious):
spur = generate_spurious_annotation(existing_bboxes, rng)
spur["id"] = next_id + i
corrupted.append(spur)
existing_bboxes.append(spur["bbox"])
log.append(f"Added spurious ann {spur['id']}")
return corrupted, log
|