File size: 17,108 Bytes
505117b | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 | from __future__ import annotations
import argparse
import json
import re
from collections import defaultdict
from pathlib import Path
import cv2
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
from pycocotools import mask as mask_utils
from pycocotools.coco import COCO
FRAME_PATTERN = re.compile(r"S(?P<sequence>\d+)", re.IGNORECASE)
COVERAGE_REGIME_BINS = [-np.inf, 0.05, 0.15, np.inf]
COVERAGE_REGIME_LABELS = ["low (<=0.05)", "intermediate (0.05-0.15)", "high (>0.15)"]
def main() -> None:
parser = argparse.ArgumentParser(
description="Audit BubbleID-Flow holdout robustness from archived COCO predictions."
)
parser.add_argument("--train-json", required=True)
parser.add_argument("--val-json", required=True)
parser.add_argument("--predictions-json", required=True)
parser.add_argument("--output-dir", required=True)
parser.add_argument("--bootstrap-samples", type=int, default=10000)
parser.add_argument("--seed", type=int, default=20260812)
parser.add_argument(
"--score-thresholds", default="0.05,0.10,0.20,0.30,0.40,0.50,0.70"
)
parser.add_argument("--detection-caps", default="50,100,200,300")
args = parser.parse_args()
output_dir = Path(args.output_dir)
output_dir.mkdir(parents=True, exist_ok=True)
thresholds = [float(value) for value in args.score_thresholds.split(",")]
caps = [int(value) for value in args.detection_caps.split(",")]
if not thresholds or not caps or min(caps) <= 0:
raise ValueError("At least one score threshold and positive detection cap are required")
coco = COCO(args.val_json)
predictions = json.loads(Path(args.predictions_json).read_text(encoding="utf-8"))
predictions_by_image: dict[int, list[dict]] = defaultdict(list)
for prediction in predictions:
predictions_by_image[int(prediction["image_id"])].append(prediction)
for image_predictions in predictions_by_image.values():
image_predictions.sort(key=lambda item: float(item["score"]), reverse=True)
rows_by_setting = []
selected_rows = None
for threshold in thresholds:
for cap in caps:
rows = evaluate_setting(coco, predictions_by_image, threshold, cap)
summary = summarize_metrics(rows)
summary.update(
{
"score_threshold": threshold,
"detection_cap": cap,
"images_at_cap": int((rows["predicted_instances"] == cap).sum()),
}
)
rows_by_setting.append(summary)
if np.isclose(threshold, 0.30) and cap == 300:
selected_rows = rows
if selected_rows is None:
raise ValueError("The required manuscript operating point (threshold 0.30, cap 300) is missing")
sensitivity = pd.DataFrame(rows_by_setting).sort_values(
["detection_cap", "score_threshold"]
)
sensitivity.to_csv(output_dir / "segmentation_operating_point_sensitivity.csv", index=False)
split_audit, split_summary = audit_split_similarity(args.train_json, args.val_json)
selected_rows = selected_rows.merge(split_audit, on=["image_id", "file_name"], how="left")
selected_rows.to_csv(output_dir / "segmentation_metrics_by_image.csv", index=False)
image_groups = build_coverage_group_manifest(coco, selected_rows)
image_groups.to_csv(output_dir / "segmentation_evaluation_image_groups.csv", index=False)
regime = summarize_by_regime(selected_rows)
regime.to_csv(output_dir / "segmentation_metrics_by_coverage_regime.csv", index=False)
bootstrap = bootstrap_intervals(
selected_rows, samples=args.bootstrap_samples, seed=args.seed
)
operating_point = summarize_metrics(selected_rows)
robustness_summary = {
"operating_point": {
"score_threshold": 0.30,
"detection_cap": 300,
**operating_point,
},
"bootstrap": {
"samples": args.bootstrap_samples,
"seed": args.seed,
"confidence_level": 0.95,
"interval_method": "image-level percentile bootstrap",
"metrics": bootstrap,
},
"split_audit": split_summary,
"interpretation": {
"supported": "internal same-sequence reconstruction of projected vapor area",
"not_supported": "experiment-held-out generalization or individual-bubble statistics",
},
}
(output_dir / "segmentation_robustness_summary.json").write_text(
json.dumps(robustness_summary, indent=2, sort_keys=True), encoding="utf-8"
)
make_figure(selected_rows, sensitivity, regime, output_dir)
print(json.dumps(robustness_summary, indent=2, sort_keys=True))
def evaluate_setting(
coco: COCO,
predictions_by_image: dict[int, list[dict]],
threshold: float,
cap: int,
) -> pd.DataFrame:
rows = []
for image_id in sorted(coco.getImgIds()):
record = coco.loadImgs([image_id])[0]
height, width = int(record["height"]), int(record["width"])
ground_truth = np.zeros((height, width), dtype=bool)
for annotation in coco.loadAnns(coco.getAnnIds(imgIds=[image_id])):
ground_truth |= coco.annToMask(annotation).astype(bool)
retained = [
item
for item in predictions_by_image.get(image_id, [])
if float(item["score"]) >= threshold
][:cap]
predicted = np.zeros((height, width), dtype=bool)
for item in retained:
predicted |= mask_utils.decode(item["segmentation"]).astype(bool)
intersection = int(np.count_nonzero(ground_truth & predicted))
union = int(np.count_nonzero(ground_truth | predicted))
ground_truth_area = int(np.count_nonzero(ground_truth))
predicted_area = int(np.count_nonzero(predicted))
pixel_count = ground_truth.size
denominator = ground_truth_area + predicted_area
rows.append(
{
"image_id": image_id,
"file_name": Path(record["file_name"]).name,
"ground_truth_vapor_area_fraction": ground_truth_area / pixel_count,
"predicted_vapor_area_fraction": predicted_area / pixel_count,
"signed_area_fraction_error": (predicted_area - ground_truth_area)
/ pixel_count,
"absolute_area_fraction_error": abs(predicted_area - ground_truth_area)
/ pixel_count,
"union_mask_iou": intersection / union if union else 1.0,
"union_mask_dice": 2 * intersection / denominator if denominator else 1.0,
"predicted_instances": len(retained),
}
)
return pd.DataFrame(rows)
def summarize_metrics(frame: pd.DataFrame) -> dict[str, float | int]:
return {
"images": int(len(frame)),
"mean_union_mask_iou": float(frame["union_mask_iou"].mean()),
"median_union_mask_iou": float(frame["union_mask_iou"].median()),
"mean_union_mask_dice": float(frame["union_mask_dice"].mean()),
"mean_absolute_area_fraction_error": float(
frame["absolute_area_fraction_error"].mean()
),
"mean_signed_area_fraction_error": float(
frame["signed_area_fraction_error"].mean()
),
"max_absolute_area_fraction_error": float(
frame["absolute_area_fraction_error"].max()
),
"max_predicted_instances": int(frame["predicted_instances"].max()),
}
def bootstrap_intervals(
frame: pd.DataFrame, *, samples: int, seed: int
) -> dict[str, dict[str, float]]:
if samples <= 0:
raise ValueError("bootstrap-samples must be positive")
metric_columns = {
"mean_union_mask_iou": "union_mask_iou",
"mean_union_mask_dice": "union_mask_dice",
"mean_absolute_area_fraction_error": "absolute_area_fraction_error",
"mean_signed_area_fraction_error": "signed_area_fraction_error",
}
rng = np.random.default_rng(seed)
indices = rng.integers(0, len(frame), size=(samples, len(frame)))
output = {}
for name, column in metric_columns.items():
values = frame[column].to_numpy(dtype=float)
samples_mean = values[indices].mean(axis=1)
low, high = np.quantile(samples_mean, [0.025, 0.975])
output[name] = {
"estimate": float(values.mean()),
"ci95_low": float(low),
"ci95_high": float(high),
}
return output
def summarize_by_regime(frame: pd.DataFrame) -> pd.DataFrame:
grouped = frame.assign(
coverage_regime=coverage_regimes(frame["ground_truth_vapor_area_fraction"])
).groupby("coverage_regime", observed=True)
rows = []
for regime, group in grouped:
rows.append({"coverage_regime": str(regime), **summarize_metrics(group)})
return pd.DataFrame(rows)
def coverage_regimes(values: pd.Series) -> pd.Categorical:
return pd.cut(
values,
bins=COVERAGE_REGIME_BINS,
labels=COVERAGE_REGIME_LABELS,
)
def build_coverage_group_manifest(coco: COCO, frame: pd.DataFrame) -> pd.DataFrame:
"""Attach original COCO paths and manual-coverage groups to Fig. 5 inputs."""
records = pd.DataFrame(
{
"image_id": [image_id for image_id in sorted(coco.getImgIds())],
"coco_image_path": [
(
Path(Path(coco.loadImgs([image_id])[0]["file_name"]).parent.name)
/ Path(coco.loadImgs([image_id])[0]["file_name"]).name
).as_posix()
for image_id in sorted(coco.getImgIds())
],
}
)
columns = [
"image_id",
"file_name",
"ground_truth_vapor_area_fraction",
"predicted_vapor_area_fraction",
"union_mask_iou",
"union_mask_dice",
"absolute_area_fraction_error",
"predicted_instances",
"source_sequence_index",
"same_nominal_frame_in_training",
"nearest_training_frame_distance",
"most_similar_training_file",
]
manifest = frame[columns].merge(records, on="image_id", how="left", validate="one_to_one")
manifest.insert(
2,
"coverage_regime",
coverage_regimes(manifest["ground_truth_vapor_area_fraction"]),
)
return manifest.sort_values(
["coverage_regime", "ground_truth_vapor_area_fraction", "file_name"],
kind="stable",
).reset_index(drop=True)
def audit_split_similarity(train_json: str, val_json: str) -> tuple[pd.DataFrame, dict]:
train_records = json.loads(Path(train_json).read_text(encoding="utf-8"))["images"]
val_records = json.loads(Path(val_json).read_text(encoding="utf-8"))["images"]
train_images = [(record, load_gray(record["file_name"])) for record in train_records]
train_sequence = [frame_sequence(record["file_name"]) for record, _ in train_images]
rows = []
for record in val_records:
val_image = load_gray(record["file_name"])
val_sequence = frame_sequence(record["file_name"])
best = None
for (train_record, train_image), sequence in zip(train_images, train_sequence):
correlation, normalized_mae = image_similarity(val_image, train_image)
candidate = (correlation, -normalized_mae, train_record, sequence, normalized_mae)
if best is None or candidate[:2] > best[:2]:
best = candidate
assert best is not None
correlation, _, nearest_record, nearest_sequence, normalized_mae = best
same_nominal = any(sequence == val_sequence for sequence in train_sequence)
min_distance = min(abs(sequence - val_sequence) for sequence in train_sequence)
rows.append(
{
"image_id": int(record["id"]),
"file_name": Path(record["file_name"]).name,
"source_sequence_index": val_sequence,
"same_nominal_frame_in_training": bool(same_nominal),
"nearest_training_frame_distance": int(min_distance),
"most_similar_training_file": Path(nearest_record["file_name"]).name,
"nearest_training_image_correlation": float(correlation),
"nearest_training_normalized_mae": float(normalized_mae),
}
)
frame = pd.DataFrame(rows)
summary = {
"train_images": len(train_records),
"holdout_images": len(val_records),
"distinct_camera_prefixes": 1,
"holdout_with_same_nominal_frame_in_training": int(
frame["same_nominal_frame_in_training"].sum()
),
"holdout_with_training_frame_within_one_index": int(
(frame["nearest_training_frame_distance"] <= 1).sum()
),
"median_nearest_training_frame_distance": float(
frame["nearest_training_frame_distance"].median()
),
"median_nearest_training_image_correlation": float(
frame["nearest_training_image_correlation"].median()
),
"split_classification": "same-sequence internal holdout",
}
return frame, summary
def frame_sequence(path: str) -> int:
match = FRAME_PATTERN.search(Path(path).stem)
if not match:
raise ValueError(f"Could not parse source sequence from {path}")
return int(match.group("sequence"))
def load_gray(path: str) -> np.ndarray:
image = cv2.imread(str(path), cv2.IMREAD_GRAYSCALE)
if image is None:
raise ValueError(f"Could not read image {path}")
return cv2.resize(image, (256, 32), interpolation=cv2.INTER_AREA).astype(np.float32)
def image_similarity(first: np.ndarray, second: np.ndarray) -> tuple[float, float]:
first_flat = first.ravel()
second_flat = second.ravel()
if np.std(first_flat) == 0 or np.std(second_flat) == 0:
correlation = float(first_flat.mean() == second_flat.mean())
else:
correlation = float(np.corrcoef(first_flat, second_flat)[0, 1])
normalized_mae = float(np.mean(np.abs(first - second)) / 255.0)
return correlation, normalized_mae
def make_figure(
frame: pd.DataFrame,
sensitivity: pd.DataFrame,
regime: pd.DataFrame,
output_dir: Path,
) -> None:
plt.rcParams.update(
{
"font.family": "Arial",
"font.size": 10,
"axes.labelsize": 10,
"xtick.labelsize": 9,
"ytick.labelsize": 9,
"legend.fontsize": 9,
"axes.linewidth": 0.8,
"mathtext.fontset": "custom",
"mathtext.rm": "Arial",
"mathtext.it": "Arial:italic",
"mathtext.bf": "Arial:bold",
"savefig.dpi": 300,
}
)
fig, axes = plt.subplots(2, 2, figsize=(7.2, 5.6), constrained_layout=True)
blue, orange, green = "#0072B2", "#D55E00", "#009E73"
ax = axes[0, 0]
x = frame["ground_truth_vapor_area_fraction"]
y = frame["predicted_vapor_area_fraction"]
limit = max(float(x.max()), float(y.max())) * 1.08
ax.scatter(x, y, color=blue, edgecolor="white", linewidth=0.5, s=28)
ax.plot([0, limit], [0, limit], color="black", linewidth=1, linestyle="--")
ax.set(xlabel="Manual projected vapor coverage", ylabel="Predicted projected vapor coverage", xlim=(0, limit), ylim=(0, limit))
ax = axes[0, 1]
ax.scatter(x, frame["union_mask_iou"], color=green, edgecolor="white", linewidth=0.5, s=28)
ax.set(xlabel="Manual projected vapor coverage", ylabel="Union-mask IoU", ylim=(0, 1))
cap300 = sensitivity[sensitivity["detection_cap"] == 300]
ax = axes[1, 0]
ax.plot(cap300["score_threshold"], cap300["mean_union_mask_iou"], marker="o", color=blue, linestyle="--", label="Mean IoU")
ax.plot(cap300["score_threshold"], cap300["mean_union_mask_dice"], marker="s", color=green, linestyle="--", label="Mean Dice")
ax.axvline(0.30, color="black", linestyle="--", linewidth=0.9)
ax.set(xlabel="Score threshold", ylabel="Overlap metric", ylim=(0, 1))
ax.legend(frameon=False, loc="lower right")
ax = axes[1, 1]
ax.plot(cap300["score_threshold"], cap300["mean_absolute_area_fraction_error"], marker="o", color=orange, linestyle="--", label="MAE")
ax.plot(cap300["score_threshold"], np.abs(cap300["mean_signed_area_fraction_error"]), marker="s", color=blue, linestyle="--", label="Absolute bias")
ax.axvline(0.30, color="black", linestyle="--", linewidth=0.9)
ax.set(xlabel="Score threshold", ylabel="Projected-area-fraction error", ylim=(0, None))
ax.legend(
frameon=False,
loc="upper center",
bbox_to_anchor=(0.5, -0.28),
ncol=2,
)
for label, ax in zip("abcd", axes.ravel()):
ax.text(0.0, 1.04, f"({label})", transform=ax.transAxes, va="bottom", clip_on=False)
ax.tick_params(direction="in", top=True, right=True)
ax.grid(False)
for suffix in ("pdf", "png"):
fig.savefig(output_dir / f"Figure_5_segmentation_robustness.{suffix}", bbox_inches="tight")
plt.close(fig)
if __name__ == "__main__":
main()
|