vnguyen/product-recognition-demo
1
0
Fork 1
Code Issues Pull requests 1 Projects Releases Packages Wiki Activity Actions

fix(catalog): re-source sku_l09 and enforce the pyramid floor

Browse source 
sku_l09.jpg was 407x305, so its 400 px pyramid level was a 305->400 upscale: a
blur with no high-frequency detail to disagree with, which won 31 of 54 crops
on the shelf it was cut from. Re-sourced at 1000x1000.

The floor is now a hard gate in rebuild_catalog.py - a view under 800 px tall
is refused, not reported - because reporting it is what let the 305 px view
ship. data/packshots/source/sku_l07.jpg (600x600) is refused on this rule.

README records the real sizes and the provenance for every row.
This commit is contained in:
kiyo9w 2026-09-29 09:28:16 +09:00
parent 5094b5f6ca
commit 730a5e782b
4 changed files with 274 additions and 102 deletions
Download patch file Download diff file Expand all files Collapse all files

62
recognization/code/rebuild_catalog.py
View file

@ -7,7 +7,9 @@ its own exam.
This rebuilds a catalog directory from named packshots and refuses, per view:
* a view shorter than MIN_HEIGHT or narrower than MIN_WIDTH,
* a view shorter than MIN_HEIGHT, narrower than MIN_WIDTH, or shorter than
PYRAMID_FLOOR (a source under it cannot be stored as the protocol's
200/400/800 px pyramid, so its scores would not be comparable with the rest),
* a view whose capture session appears in --eval-manifest,
* a packshot whose catalog name is not supported by its OCR text,
* a view that is a near-duplicate of another view of the same SKU,
@ -63,10 +65,16 @@ MIN_WIDTH = 120
# The enrollment protocol's floor: a reference view under 160 px tall carries no
# text geometry (facings under 60 px wide failed every measured lane).
MIN_HEIGHT = 160
# Correct packshots measured 0.79-1.00 of the catalog name's bigrams in their
# OCR text; the wrong-size Kewpie 450g packshot tops out at 0.64 against every
# name in the catalog. 0.70 sits between them with room for OCR noise.
MIN_NAME_OCR = 0.70
# The name/OCR gate, re-measured on the shipped packshots with the strict name
# side in place, each image read four ways and scored against all 12 catalog
# names: correct packshots 0.9167-1.0000 (seven read their whole name; l11's
# read omits "マヨネーズ" and is the floor), the wrong-product packshot in
# references/rejected/ 0.0476-0.5833. 0.75 is the midpoint: 0.167 of headroom
# above every correct packshot and 0.167 below the wrong-product ceiling, so
# OCR noise on either side does not move a verdict. Siblings are not this
# gate's job - the コクうま packshot scores 0.85 against the マヨネーズ name,
# which is what text_discriminated_siblings flags, not a gap a threshold closes.
MIN_NAME_OCR = 0.75
# Vision drops small print at native scale: the l11 packshot reads ハーフ only
# when upscaled (0.57 -> 0.79). Reading at 2x is what makes the check about the
# package rather than about Vision's recall.
@ -75,10 +83,12 @@ OCR_SCALE = 2.0
DUPLICATE_COSINE = 0.98
# The enrollment protocol's pyramid target: a source at least this tall can be
# stored as a 200/400/800 px pyramid. Below it the view still enrolls - the hard
# reject is MIN_HEIGHT - but its scores are not comparable with the rest, which
# is the documented l09 defect: at 305 px its inlier scores dwarf every other
# reference (85-129 against 10-40) and it wins facings it should not.
# stored as a 200/400/800 px pyramid, so a view under it cannot be scored
# against the rest. Refused, not reported: the l09 defect was a 305 px view
# whose inlier scores (85-129) dwarfed every other reference (10-40) and it won
# facings it should not, and a report is exactly the kind of thing that gets
# read once and then ignored. The fix is re-sourcing, which is a data task
# this tool should not let through the door in the first place.
PYRAMID_FLOOR = 800
# Official pack-shot pages publish a single front view per JAN, so a catalog
@ -317,6 +327,12 @@ def audit(
refused.append({"sku": sku, "view": name,
"reason": f"{image.shape[1]}px wide, below the {MIN_WIDTH}px floor"})
continue
if image.shape[0] < PYRAMID_FLOOR:
refused.append({"sku": sku, "view": name,
"reason": f"{image.shape[0]}px tall, below the {PYRAMID_FLOOR}px "
"pyramid floor; its scores would not be comparable "
"with the rest - re-source this view"})
continue
if ocr_engine is not None and names is not None:
try:
text = best_orientation_text(ocr_engine, image, names.get(sku, ""))
@ -453,27 +469,6 @@ def load_shelf_skus(path: Optional[str]) -> Optional[Set[str]]:
raise RebuildError(f"{path}: expected a JSON list or one SKU per line")
def below_pyramid_floor(kept: Dict[str, List[str]], source: str) -> List[Dict[str, Any]]:
"""Enrolled views too small to store as the protocol's pyramid.
Reported, not refused: the view is usable and dropping it would remove the
SKU from the catalog entirely, which is worse than an incomparable score.
The fix is re-sourcing the packshot, which is a data task.
"""
small: List[Dict[str, Any]] = []
for sku, filenames in sorted(kept.items()):
for filename in filenames:
image = cv2.imread(os.path.join(source, filename))
if image is not None and image.shape[0] < PYRAMID_FLOOR:
small.append({
"sku": sku,
"view": filename,
"height": int(image.shape[0]),
"floor": PYRAMID_FLOOR,
})
return small
def write_catalog(kept: Dict[str, List[str]], source: str, output: str, csv_path: Optional[str]) -> None:
"""Copy the kept views and the whole CSV into a staging dir, then swap it in."""
staging = tempfile.mkdtemp(prefix=".catalog-", dir=os.path.dirname(os.path.abspath(output)) or ".")
@ -582,7 +577,6 @@ def main(argv: Optional[Sequence[str]] = None) -> int:
# and vetoes it from the CSV alone, so it is a coverage gap, not a refusal.
text_only = sorted(set(names) - set(kept))
siblings = text_discriminated_siblings(kept, args.source, names)
small = below_pyramid_floor(kept, args.source)
no_source = unattributed(args.source)
verdict = live_ready_verdict(kept, names, shelf_skus)
report = {
@ -596,7 +590,6 @@ def main(argv: Optional[Sequence[str]] = None) -> int:
"ocr": engine.name if ocr_engine is not None else "skipped",
"text_only_skus": text_only,
"text_discriminated_siblings": siblings,
"below_pyramid_floor": small,
"unattributed_skus": no_source,
"completeness": verdict,
"skus": {sku: views for sku, views in sorted(kept.items())},
@ -616,11 +609,6 @@ def main(argv: Optional[Sequence[str]] = None) -> int:
for row in siblings:
print(f" {row['sku']:<10} {row['view']:<20} ~ {row['sibling_sku']} "
f"({row['sibling_view']}) cosine {row['cosine']}")
if small:
print(f"\nbelow the {PYRAMID_FLOOR}px pyramid floor ({len(small)}):")
for row in small:
print(f" {row['sku']:<10} {row['view']:<20} {row['height']}px tall - "
"scores are not comparable with the rest; re-source this view")
if no_source:
print(f"\nno recorded source ({len(no_source)}): {', '.join(no_source)} - "
"usable, but provenance is missing; re-sourcing would close it")

144
recognization/code/test_rebuild_catalog.py
View file

@ -1,11 +1,13 @@
"""
Contract tests for the enrollment tool's reporting checks.
Contract tests for the enrollment tool's gates.
The refusal rules (size floors, session leakage, name/OCR, duplicates) are what
the tool exists for; these tests cover the two checks that report rather than
refuse, because both are properties of the catalog as a whole and neither is
visible from a single view: cross-SKU near-duplicates, and the completeness
rule that decides whether a bay is live-ready.
The refusal rules (size floors including the pyramid floor, session leakage,
name/OCR, duplicates) are what the tool exists for, and they are pinned here
against the measured distribution rather than against invented numbers. The
two checks that report rather than refuse are covered too, because both are
properties of the catalog as a whole and neither is visible from a single
view: cross-SKU near-duplicates, and the completeness rule that decides whether
a bay is live-ready.
"""
from __future__ import annotations
@ -23,11 +25,13 @@ import cv2 # noqa: E402
import numpy as np # noqa: E402
from rebuild_catalog import ( # noqa: E402
MIN_NAME_OCR,
PYRAMID_FLOOR,
RebuildError,
below_pyramid_floor,
audit,
live_ready_verdict,
load_shelf_skus,
name_ocr_similarity,
text_discriminated_siblings,
)
@ -75,8 +79,13 @@ class SiblingFlagTests(unittest.TestCase):
text_discriminated_siblings({"sku_x1": ["sku_x1.jpg", "sku_x1_2.jpg"]}, root), []
)
class PyramidFloorTests(unittest.TestCase):
"""The l09 defect: a view too small to store as the protocol's pyramid."""
"""The l09 defect: a 305 px view whose scores dwarf every other reference.
Refused, not reported. Reporting it is what let the defect ship - the view
enrolled anyway, because a line in a catalog README is not a gate.
"""
def make_source(self, height):
root = tempfile.mkdtemp(prefix="floor-")
@ -84,17 +93,112 @@ class PyramidFloorTests(unittest.TestCase):
cv2.imwrite(os.path.join(root, "sku_x1.jpg"), image)
return root
def test_a_small_view_is_reported(self):
root = self.make_source(305)
small = below_pyramid_floor({"sku_x1": ["sku_x1.jpg"]}, root)
self.assertEqual(len(small), 1, small)
self.assertEqual(small[0]["sku"], "sku_x1")
self.assertEqual(small[0]["height"], 305)
self.assertEqual(small[0]["floor"], PYRAMID_FLOOR)
def test_a_view_under_the_floor_is_refused(self):
kept, refused = audit(self.make_source(305), min_views=1)
self.assertEqual(kept, {}, refused)
self.assertEqual(len(refused), 1, refused)
self.assertIn("305px", refused[0]["reason"])
self.assertIn("pyramid floor", refused[0]["reason"])
def test_a_full_size_view_is_not_reported(self):
root = self.make_source(1000)
self.assertEqual(below_pyramid_floor({"sku_x1": ["sku_x1.jpg"]}, root), [])
def test_the_floor_is_exact(self):
"""799 refused, 800 kept: the boundary is the number, not a rounding."""
under = audit(self.make_source(PYRAMID_FLOOR - 1), min_views=1)
self.assertEqual(under[0], {}, under[1])
at = audit(self.make_source(PYRAMID_FLOOR), min_views=1)
self.assertEqual(at[0], {"sku_x1": ["sku_x1.jpg"]}, at[1])
self.assertEqual(at[1], [])
def test_a_full_size_view_still_enrolls(self):
kept, refused = audit(self.make_source(1000), min_views=1)
self.assertEqual(kept, {"sku_x1": ["sku_x1.jpg"]}, refused)
self.assertEqual(refused, [])
class NameOcrTests(unittest.TestCase):
"""The wrong-product gate: a packshot's OCR must explain its catalog row.
Pinned to the measured distribution, not to invented numbers. The reads are
what the vision backend returned for the shipped packshots and for the
rejected Kewpie 450g; recording them keeps the test on the pure function
instead of on Vision's recall, which moves with the macOS version.
"""
# Worst correct packshot in stock-lopia-clean: the l11 read carries
# "キューピー ハーフ" and "400g" but never "マヨネーズ".
WORST_CORRECT = (
"キューピー ハーフマヨネーズ 400g",
"2 キューピーハーフ New！ 卵黄ならではのコクとうま味 "
"※写真はイメージです。 カロリー 50%ット 当社マヨネーズ比 400g",
0.9167,
)
# The recorded l09/l10 mix-up: Kewpie Mayonnaise 450g, filed under an
# Ajinomoto row. This is its best score against any of the 12 catalog names.
WRONG_PRODUCT = (
"キューピー からしマヨネーズ 200g",
"キューピーマヨネーズ KEWPIE MAYONNAISE ® 450g",
0.5833,
)
def test_a_correct_packshot_clears_the_gate(self):
name, text, measured = self.WORST_CORRECT
self.assertAlmostEqual(name_ocr_similarity(name, text), measured, places=3)
self.assertGreaterEqual(name_ocr_similarity(name, text), MIN_NAME_OCR)
def test_the_wrong_size_packshot_is_refused(self):
"""The check that exists to stop the l09/l10 mix-up keeps its teeth."""
name, text, measured = self.WRONG_PRODUCT
self.assertAlmostEqual(name_ocr_similarity(name, text), measured, places=3)
self.assertLess(name_ocr_similarity(name, text), MIN_NAME_OCR)
def test_the_threshold_sits_between_the_two_measured_bands(self):
"""The recalibration pin. Lower it and the wrong product enrolls."""
worst_correct = name_ocr_similarity(*self.WORST_CORRECT[:2])
best_wrong = name_ocr_similarity(*self.WRONG_PRODUCT[:2])
self.assertGreater(worst_correct, best_wrong, "the bands overlap; no threshold separates them")
self.assertGreaterEqual(MIN_NAME_OCR, best_wrong)
self.assertLessEqual(MIN_NAME_OCR, worst_correct)
self.assertGreaterEqual(MIN_NAME_OCR - best_wrong, 0.1, "no headroom below the wrong-product ceiling")
self.assertGreaterEqual(worst_correct - MIN_NAME_OCR, 0.1, "no headroom above the worst correct packshot")
def test_a_maker_prefix_gap_does_not_admit_a_wrong_row(self):
"""l07: its product tokens all read, its maker and size do not.
The catalog names l07 "丸和油脂 ラッキーマヨネーズ 200g" and the packshot
reads "ラッキー マヨネーズ 調味料（アミノ酸）不使用" - every product bigram,
none of the maker or the size. The 0.5714 it scores is inside the
wrong-product band and stays refused: the gap is a name-token problem,
not a threshold to move. Scoping the check to the product tokens would
admit it; that is a change to this gate, not a reason to lower it.
"""
score = name_ocr_similarity("丸和油脂 ラッキーマヨネーズ 200g",
"ラッキー マヨネーズ 調味料（アミノ酸）不使用")
self.assertAlmostEqual(score, 0.5714, places=3)
self.assertLess(score, MIN_NAME_OCR)
def test_audit_refuses_a_packshot_the_name_never_supports(self):
"""l04 at 0.00: the upright read is three stray characters."""
root = tempfile.mkdtemp(prefix="name-ocr-")
cv2.imwrite(os.path.join(root, "sku_x1.jpg"),
cv2.resize(artwork(), (300, 800), interpolation=cv2.INTER_NEAREST))
class FixedOcr:
name = "fixed"
def read(self, image):
class Line:
text = "明太手屋さんが つくつ た コ"
class Read:
lines = [Line()]
return Read()
kept, refused = audit(root, min_views=1,
names={"sku_x1": "やまや めんたいマヨネーズタイプ 200g"},
ocr_engine=FixedOcr())
self.assertEqual(kept, {}, refused)
self.assertEqual(len(refused), 1, refused)
self.assertIn("name/OCR", refused[0]["reason"])
class CompletenessVerdictTests(unittest.TestCase):
@ -257,7 +361,7 @@ class OrientationTests(unittest.TestCase):
root = tempfile.mkdtemp(prefix="orient-")
# Wider than tall, so the upright pass is the "wide" read (noise) and
# only a 90-degree rotation produces the "tall" read that matches.
cv2.imwrite(os.path.join(root, "sku_x1.jpg"), self.marked(300, 400))
cv2.imwrite(os.path.join(root, "sku_x1.jpg"), self.marked(800, 1000))
engine = self.FakeOcr({"wide": "つくつ|た", "tall": "キユーピー|エッグケア|305g"})
kept, refused = audit(root, min_views=1, names={"sku_x1": "キューピー エッグケア 305g"},
ocr_engine=engine)
@ -269,7 +373,7 @@ class OrientationTests(unittest.TestCase):
from rebuild_catalog import audit
root = tempfile.mkdtemp(prefix="orient-bad-")
cv2.imwrite(os.path.join(root, "sku_x1.jpg"), self.marked(400, 300))
cv2.imwrite(os.path.join(root, "sku_x1.jpg"), self.marked(800, 300))
engine = self.FakeOcr({"wide": "キューピー|マヨネーズ", "tall": "キューピー|マヨネーズ"})
kept, refused = audit(root, min_views=1, names={"sku_x1": "やまや めんたいマヨネーズタイプ 200g"},
ocr_engine=engine)