product-recognition-demo/recognization/code
tngo b6c7c4b667 Score package text against the row's name, not its packshot's OCR
The text lane built each row's signature from its printed name plus the OCR of
its enrolled packshot, and scored a crop against that whole string. l09's
packshot prints "マヨネーズ", which is l10's actual name, so l09 could explain
an l10 crop's text better than l10's own name could. Every l10 facing reads
"ピュア セレクト マヨネーズ" with no "コクうま" and no "65", and l09 still won
or tied them.

The signature is now the row's name alone, weighted by inverse document
frequency over the catalog so copy every sibling prints counts for little and
the words that name one product dominate. The packshot OCR is gone: every row
has a name, so it added nothing but the confusion.

Measured on image6, local lane at live resolution:

  l10 correct    2/11 -> 5/11
  wrong          3 -> 1
  precision      95% -> 97%
  l09 hub share  34% (alarm) -> 29% (clear)

Full resolution end-to-end 59% at 88% precision, live 56% at 97%.
2026-09-29 00:37:56 +09:00
..
web fix: restore reliable authenticated live recognition 2026-09-28 12:38:15 +09:00
__init__.py
app.py feat(identity): local-feature matcher plus text lane as the default backend 2026-09-28 22:27:21 +09:00
bench_engine.py
bench_identity.py feat: a multi-view embedding gallery, because one template per SKU cannot work 2026-09-24 23:43:27 +09:00
build_store_from_shelf.py fix: require explicit safe shelf enrollment paths 2026-09-28 13:41:43 +09:00
collect_shelf_images.py feat: a live path that uses the enrolled store, and a demo server a phone can reach 2026-09-25 10:51:57 +09:00
detector.py
empty_shelf_detector.py
engine.py perf: read each package once per track, not once per attempt 2026-09-28 23:25:09 +09:00
eval_external_images.py test: evaluate on 54 real store photos the pipeline has never seen 2026-09-25 02:21:00 +09:00
eval_identity.py feat: refuse a label file whose rows carry no provenance 2026-09-28 23:31:53 +09:00
eval_live.py docs: name the local lane in the evaluators' own lane vocabulary 2026-09-28 22:34:06 +09:00
feature_matcher.py
finetune_identity.py fix: refuse unavailable gallery and pin head trainer 2026-09-28 14:41:57 +09:00
gallery.py feat(identity): local-feature matcher plus text lane as the default backend 2026-09-28 22:27:21 +09:00
identity.py feat(identity): local-feature matcher plus text lane as the default backend 2026-09-28 22:27:21 +09:00
identity_lanes.py Score package text against the row's name, not its packshot's OCR 2026-09-29 00:37:56 +09:00
inventory.py
live.py Fix label frame mapping and reference upscaling; add the stream G1 harness 2026-09-29 00:30:17 +09:00
ocr.py feat: add bounded crop evidence and keyed gallery cache 2026-09-28 15:49:34 +09:00
rebuild_catalog.py fix: read --shelf-tags before the catalog swap 2026-09-28 23:03:08 +09:00
reconcile.py fix: a tag may not accuse on weak evidence, nor name a distant product 2026-09-25 01:30:03 +09:00
record.py fix: record and require the walk identity backend 2026-09-28 14:42:03 +09:00
score_stream.py Fix label frame mapping and reference upscaling; add the stream G1 harness 2026-09-29 00:30:17 +09:00
shelf_map.py feat: a shelf map, so a facing outlives the track id that found it 2026-09-25 01:54:38 +09:00
stream_client.py Fix label frame mapping and reference upscaling; add the stream G1 harness 2026-09-29 00:30:17 +09:00
tag_cards.py fix: restore reliable authenticated live recognition 2026-09-28 12:38:15 +09:00
tags.py fix: bind tag projection to each frame origin 2026-09-28 14:41:49 +09:00
test_api.py feat(identity): local-feature matcher plus text lane as the default backend 2026-09-28 22:27:21 +09:00
test_engine.py Score package text against the row's name, not its packshot's OCR 2026-09-29 00:37:56 +09:00
test_eval_labels.py feat: refuse a label file whose rows carry no provenance 2026-09-28 23:31:53 +09:00
test_eval_live.py fix: record and require the walk identity backend 2026-09-28 14:42:03 +09:00
test_gallery.py fix: serialize gallery inference so Metal cannot be entered twice 2026-09-28 17:22:55 +09:00
test_identity.py refactor: delete the legacy session model 2026-09-28 17:47:50 +09:00
test_live.py perf: read each package once per track, not once per attempt 2026-09-28 23:25:09 +09:00
test_ocr_tags.py fix: bind tag projection to each frame origin 2026-09-28 14:41:49 +09:00
test_real_model.py docs: API contract, README for the server and live mode, real-model test 2026-09-24 12:46:54 +09:00
test_rebuild_catalog.py fix: read --shelf-tags before the catalog swap 2026-09-28 23:03:08 +09:00
test_reconcile.py fix: a tag may not accuse on weak evidence, nor name a distant product 2026-09-25 01:30:03 +09:00
test_record.py fix: record and require the walk identity backend 2026-09-28 14:42:03 +09:00
test_shelf_map.py feat: a shelf map, so a facing outlives the track id that found it 2026-09-25 01:54:38 +09:00
test_tag_cards.py fix: the number on the tag is the regular price, and a price must be corroborated 2026-09-25 01:29:46 +09:00
test_tag_identity.py fix: a tag may not accuse on weak evidence, nor name a distant product 2026-09-25 01:30:03 +09:00
test_tracking.py
tracking.py fix: restore reliable authenticated live recognition 2026-09-28 12:38:15 +09:00
train.py
transport.py fix: guard reserved sessions and bounded response state 2026-09-28 13:41:35 +09:00