Datasets:
File size: 10,794 Bytes
7825400 | 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 | """Taxonomy-relaxed metrics driven by a frozen semantic-pair policy."""
from __future__ import annotations
from collections import defaultdict
from dataclasses import dataclass
from metrics._f_beta import ViolationCounts, aggregate_summary, summarise_counts
from metrics._labels import to_use_case_id
from metrics.localisation import _compute_iou, maximum_match_pairs
from schemas.ground_truth import GoldRecord
from schemas.inference import InferenceResponse
@dataclass(frozen=True)
class SemanticPolicy:
"""A symmetric class-pair acceptance matrix with similarity provenance."""
class_ids: tuple[str, ...]
similarities: tuple[tuple[float, ...], ...]
accepted: tuple[tuple[int, ...], ...]
@classmethod
def from_dict(cls, payload: dict) -> "SemanticPolicy":
return cls(
class_ids=tuple(payload["class_ids"]),
similarities=tuple(tuple(row) for row in payload["cosine_matrix"]),
accepted=tuple(tuple(row) for row in payload["accepted_matrix"]),
)
def _indices(self, predicted: str, gold: str) -> tuple[int, int] | None:
try:
return self.class_ids.index(predicted), self.class_ids.index(gold)
except ValueError:
return None
def allows(self, predicted: str, gold: str) -> bool:
indices = self._indices(predicted, gold)
return bool(indices and self.accepted[indices[0]][indices[1]])
def similarity(self, predicted: str, gold: str) -> float:
indices = self._indices(predicted, gold)
return self.similarities[indices[0]][indices[1]] if indices else 0.0
def _maximum_semantic_pairs(
predicted: list[str], gold: list[str], policy: SemanticPolicy
) -> list[tuple[int, int]]:
"""Maximum-cardinality deterministic matching for non-exact class pairs."""
candidates = [
sorted(
(j for j, gold_id in enumerate(gold) if pred_id != gold_id and policy.allows(pred_id, gold_id)),
key=lambda j: (-policy.similarity(pred_id, gold[j]), j),
)
for pred_id in predicted
]
gold_owner: dict[int, int] = {}
def assign(pred_index: int, seen: set[int]) -> bool:
for gold_index in candidates[pred_index]:
if gold_index in seen:
continue
seen.add(gold_index)
owner = gold_owner.get(gold_index)
if owner is None or assign(owner, seen):
gold_owner[gold_index] = pred_index
return True
return False
for pred_index in sorted(
range(len(predicted)),
key=lambda i: (-max((policy.similarity(predicted[i], gold[j]) for j in candidates[i]), default=0.0), i),
):
assign(pred_index, set())
return sorted((pred_index, gold_index) for gold_index, pred_index in gold_owner.items())
def _negative_fields(prediction_labels: set[str], record: GoldRecord, totals: dict) -> None:
if record.gold.no_violation_detected:
totals["negative_images"] += 1
if prediction_labels:
totals["false_positive_images"] += 1
else:
totals["true_negative_images"] += 1
def compute_semantic_classification_metrics(
predictions: list[InferenceResponse],
ground_truths: list[GoldRecord],
policy: SemanticPolicy,
) -> dict:
"""Score unique class sets with exact-first, one-to-one semantic matching."""
gt_by_id = {record.image_id: record for record in ground_truths}
counts: dict[str, ViolationCounts] = defaultdict(ViolationCounts)
totals = {"negative_images": 0, "true_negative_images": 0, "false_positive_images": 0}
credited = []
evaluated = 0
for prediction in predictions:
record = gt_by_id.get(prediction.image_id)
if record is None or record.gold.reported:
continue
evaluated += 1
gt_labels = {item.use_case_id for item in record.gold.violations}
pred_labels = {to_use_case_id(item.category_id) for item in prediction.detections}
exact = gt_labels & pred_labels
for label in exact:
counts[label].tp += 1
unmatched_pred = sorted(pred_labels - exact)
unmatched_gt = sorted(gt_labels - exact)
pairs = _maximum_semantic_pairs(unmatched_pred, unmatched_gt, policy)
matched_pred = {pred_index for pred_index, _ in pairs}
matched_gt = {gold_index for _, gold_index in pairs}
for pred_index, gold_index in pairs:
predicted_id, gold_id = unmatched_pred[pred_index], unmatched_gt[gold_index]
counts[gold_id].tp += 1
credited.append({
"image_id": prediction.image_id,
"predicted_class": predicted_id,
"gold_class": gold_id,
"similarity": round(policy.similarity(predicted_id, gold_id), 6),
})
for index, label in enumerate(unmatched_pred):
if index not in matched_pred:
counts[label].fp += 1
for index, label in enumerate(unmatched_gt):
if index not in matched_gt:
counts[label].fn += 1
_negative_fields(pred_labels, record, totals)
negatives = totals["negative_images"]
return {
"per_violation": summarise_counts(counts),
"summary": aggregate_summary(counts),
"n_images": evaluated,
**totals,
"negative_false_positive_rate": round(totals["false_positive_images"] / negatives, 4) if negatives else None,
"credited_matches": credited,
}
def _maximum_semantic_box_pairs(predictions: list[dict], gold: list[dict], policy: SemanticPolicy, threshold: float) -> list[tuple[int, int, float]]:
candidates = [
sorted(
(
j for j, target in enumerate(gold)
if item["label"] != target["label"]
and policy.allows(item["label"], target["label"])
and _compute_iou(item["bbox"], target["bbox"]) >= threshold
),
key=lambda j: (
-policy.similarity(item["label"], gold[j]["label"]),
-_compute_iou(item["bbox"], gold[j]["bbox"]),
gold[j]["index"],
),
)
for item in predictions
]
gold_owner: dict[int, int] = {}
def assign(pred_index: int, seen: set[int]) -> bool:
for gold_index in candidates[pred_index]:
if gold_index in seen:
continue
seen.add(gold_index)
owner = gold_owner.get(gold_index)
if owner is None or assign(owner, seen):
gold_owner[gold_index] = pred_index
return True
return False
for pred_index in sorted(range(len(predictions)), key=lambda i: predictions[i]["index"]):
assign(pred_index, set())
return sorted(
(
pred_index,
gold_index,
_compute_iou(predictions[pred_index]["bbox"], gold[gold_index]["bbox"]),
)
for gold_index, pred_index in gold_owner.items()
)
def compute_semantic_localisation_metrics(
predictions: list[InferenceResponse],
ground_truths: list[GoldRecord],
policy: SemanticPolicy,
iou_threshold: float = 0.50,
) -> dict:
"""Score boxes with exact-label matches first, then accepted class pairs."""
gt_by_id = {record.image_id: record for record in ground_truths}
counts: dict[str, ViolationCounts] = defaultdict(ViolationCounts)
totals = {"negative_images": 0, "true_negative_images": 0, "false_positive_images": 0}
credited = []
evaluated = n_predictions = n_gt_boxes = 0
for prediction in predictions:
record = gt_by_id.get(prediction.image_id)
if record is None or record.gold.reported:
continue
evaluated += 1
pred_items = [
{"index": i, "label": to_use_case_id(item.category_id), "bbox": item.bbox}
for i, item in enumerate(prediction.detections)
]
gt_items = [
{"index": i, "label": item.use_case_id, "bbox": item.bbox}
for i, item in enumerate(record.gold.violations)
]
n_predictions += len(pred_items)
n_gt_boxes += len(gt_items)
matched_pred: set[int] = set()
matched_gt: set[int] = set()
labels = {item["label"] for item in pred_items + gt_items}
for label in sorted(labels):
preds = [item for item in pred_items if item["label"] == label]
golds = [item for item in gt_items if item["label"] == label]
for local_pred, local_gt, _ in maximum_match_pairs(
[item["bbox"] for item in preds], [item["bbox"] for item in golds], iou_threshold
):
matched_pred.add(preds[local_pred]["index"])
matched_gt.add(golds[local_gt]["index"])
counts[label].tp += 1
remaining_pred = [item for item in pred_items if item["index"] not in matched_pred]
remaining_gt = [item for item in gt_items if item["index"] not in matched_gt]
semantic_pairs = _maximum_semantic_box_pairs(remaining_pred, remaining_gt, policy, iou_threshold)
semantic_pred = {pred_index for pred_index, _, _ in semantic_pairs}
semantic_gt = {gold_index for _, gold_index, _ in semantic_pairs}
for pred_index, gold_index, iou in semantic_pairs:
pred_item, gold_item = remaining_pred[pred_index], remaining_gt[gold_index]
counts[gold_item["label"]].tp += 1
credited.append({
"image_id": prediction.image_id,
"prediction_index": pred_item["index"],
"gold_index": gold_item["index"],
"predicted_class": pred_item["label"],
"gold_class": gold_item["label"],
"similarity": round(policy.similarity(pred_item["label"], gold_item["label"]), 6),
"iou": round(iou, 6),
})
for index, item in enumerate(remaining_pred):
if index not in semantic_pred:
counts[item["label"]].fp += 1
for index, item in enumerate(remaining_gt):
if index not in semantic_gt:
counts[item["label"]].fn += 1
_negative_fields({item["label"] for item in pred_items}, record, totals)
negatives = totals["negative_images"]
return {
"per_violation": summarise_counts(counts),
"summary": aggregate_summary(counts),
"iou_threshold": iou_threshold,
"n_images": evaluated,
"n_predictions": n_predictions,
"n_gt_boxes": n_gt_boxes,
**totals,
"negative_false_positive_rate": round(totals["false_positive_images"] / negatives, 4) if negatives else None,
"credited_matches": credited,
}
|