Yeva's picture
Add standalone evaluation kit
551cc83 verified
Raw
History Blame Contribute Delete
8.44 kB
from __future__ import annotations
from dataclasses import dataclass
from typing import Any
from geometry import (
Box,
polygon_area,
polygon_intersection,
point_in_convex_polygon,
)
REDUNDANT_BOX_MIN_OVERLAP_RATIO = 0.95
IMAGE_RELATED_LABELS = frozenset(
{
"Photo",
"Graphics",
"SealFigure",
"FrontPicture",
}
)
HEADER_TITLE_LIKE_LABELS = frozenset(
{
"Headline",
"Kicker",
"Banner",
"Deck",
"Subhead",
"Nameplate",
"Masthead",
"FrontStory",
}
)
IMAGE_HEADER_FILTER_LABELS = IMAGE_RELATED_LABELS | HEADER_TITLE_LIKE_LABELS
def _normalize_whitespace(text: str) -> str:
"""Local helper to avoid circular import with spatial.py."""
return " ".join(text.split())
@dataclass(frozen=True)
class Word:
box: Box
text: str
@property
def center(self) -> tuple[float, float]:
return (
(self.box.x_min + self.box.x_max) / 2.0,
(self.box.y_min + self.box.y_max) / 2.0,
)
@dataclass(frozen=True)
class AnnotationBox:
box_id: str
rect: Box
text: str
has_transcription: bool
rotation: float
polygon: tuple[tuple[float, float], ...]
bounds: Box
labels: tuple[str, ...] = ()
letter_count: int = 0
latin_or_cyrillic_letter_count: int = 0
non_armenian_letter_ratio: float = 0.0
excluded_as_non_armenian_text: bool = False
parent_box_id: str | None = None
reading_order: int | None = None
def contains_point(self, x: float, y: float) -> bool:
if not self.bounds.contains_point(x, y):
return False
return point_in_convex_polygon((x, y), self.polygon)
def overlap_area_with_box(self, box: Box) -> float:
if self.bounds.intersection(box) is None:
return 0.0
return polygon_area(polygon_intersection(box.to_polygon(), self.polygon))
@dataclass(frozen=True)
class PredictedRow:
row_id: str
words: list[Word]
@property
def text(self) -> str:
return " ".join(word.text for word in self.words if word.text.strip())
def annotation_box_contains_annotation_box(container: AnnotationBox, inner: AnnotationBox) -> bool:
if container.box_id == inner.box_id:
return False
if polygon_area(container.polygon) <= polygon_area(inner.polygon):
return False
if (
inner.bounds.x_min < container.bounds.x_min - 1e-6
or inner.bounds.y_min < container.bounds.y_min - 1e-6
or inner.bounds.x_max > container.bounds.x_max + 1e-6
or inner.bounds.y_max > container.bounds.y_max + 1e-6
):
return False
return all(container.contains_point(point[0], point[1]) for point in inner.polygon)
def annotation_box_overlap_ratio(container: AnnotationBox, inner: AnnotationBox) -> float:
inner_area = polygon_area(inner.polygon)
if inner_area <= 0:
return 0.0
overlap_area = polygon_area(polygon_intersection(inner.polygon, container.polygon))
return overlap_area / inner_area
def annotation_boxes_have_compatible_text(container: AnnotationBox, inner: AnnotationBox) -> bool:
container_text = _normalize_whitespace(container.text)
inner_text = _normalize_whitespace(inner.text)
if not container_text or not inner_text:
return False
return (
container_text == inner_text
or inner_text in container_text
or container_text in inner_text
)
def decomposable_container_box_ids(boxes: list[AnnotationBox]) -> frozenset[str]:
"""Return box_ids of containers whose text is exactly reproduced by
concatenating (in reading order) two or more boxes it contains.
Some annotations squeeze multiple side-by-side columns (e.g. a folioline
with page number / title / date, or a paragraph split across two print
columns) into one wide box, encoding the correct reading order only in
the transcription's word order. When that box also has separately
annotated children covering the same content, evaluating against the
children directly sidesteps having to guess the column layout from
geometry alone.
"""
textful = [b for b in boxes if b.has_transcription and b.text.strip()]
# `parent_id` is overloaded in this dataset: for text fragments of a
# squeezed container it means "part of this box's text", but the same
# field also links unrelated story components (headline, caption, ...)
# to a page anchor box, so it can't be trusted as a children set on its
# own — it must independently reconstruct the container's text, same as
# the geometric candidates below.
children_by_declared_parent: dict[str, list[AnnotationBox]] = {}
for b in textful:
if b.parent_box_id:
children_by_declared_parent.setdefault(b.parent_box_id, []).append(b)
def reconstructs(children: list[AnnotationBox], container_text: str) -> bool:
if len(children) < 2:
return False
order_keys = [
lambda b: (b.rect.y_min, b.rect.x_min),
lambda b: (b.rect.x_min, b.rect.y_min),
]
if all(b.reading_order is not None for b in children):
order_keys.insert(0, lambda b: b.reading_order)
for order_key in order_keys:
ordered = sorted(children, key=order_key)
reconstructed = _normalize_whitespace(" ".join(b.text for b in ordered))
if reconstructed == container_text:
return True
return False
decomposable: set[str] = set()
for container in textful:
container_text = _normalize_whitespace(container.text)
if len(container_text) < 10:
continue
# Geometric containment catches most cases, but a fraction-of-a-degree
# rotation mismatch between an annotated box and its (visually
# identical) parent can make a rotated-polygon corner fall a hair
# outside the parent, failing strict containment. `parent_id` is
# immune to that, so it's tried as an independent candidate set
# rather than merged with the geometric one (merging would pull in
# unrelated same-parent siblings and break the exact-text check).
declared_children = children_by_declared_parent.get(container.box_id, [])
geometric_children = [
b
for b in textful
if b.box_id != container.box_id
and annotation_box_contains_annotation_box(container, b)
]
if reconstructs(declared_children, container_text) or reconstructs(
geometric_children, container_text
):
decomposable.add(container.box_id)
return frozenset(decomposable)
def filter_redundant_annotation_boxes(boxes: list[AnnotationBox]) -> list[AnnotationBox]:
filtered_boxes: list[AnnotationBox] = []
for candidate in boxes:
is_redundant = False
for other in boxes:
if other.box_id == candidate.box_id:
continue
if polygon_area(other.polygon) <= polygon_area(candidate.polygon):
continue
if not annotation_boxes_have_compatible_text(other, candidate):
continue
if annotation_box_contains_annotation_box(other, candidate):
is_redundant = True
break
if annotation_box_overlap_ratio(other, candidate) >= REDUNDANT_BOX_MIN_OVERLAP_RATIO:
is_redundant = True
break
if not is_redundant:
filtered_boxes.append(candidate)
return filtered_boxes
def annotation_box_type(annotation_box: AnnotationBox) -> str | None:
return annotation_box.labels[0] if annotation_box.labels else None
def annotation_box_metadata(annotation_box: AnnotationBox) -> dict[str, Any]:
return {
"box_id": annotation_box.box_id,
"box_type": annotation_box_type(annotation_box),
"labels": list(annotation_box.labels),
}
def gt_box_report(annotation_box: AnnotationBox) -> dict[str, Any]:
# Inline rounded_box to avoid circular import with spatial.py
rounded = [round(v, 3) for v in [
annotation_box.bounds.x_min,
annotation_box.bounds.y_min,
annotation_box.bounds.x_max,
annotation_box.bounds.y_max,
]]
return {
**annotation_box_metadata(annotation_box),
"box": rounded,
}