arudaev/smart-bin-detect

Training data for Smart Bin Recognition – a validator ("is there a bin?") and an identifier ("which bin?"). The design lives in docs/04-ml-pipeline.md in the project repo, which is private; the manifests here carry per-image provenance and are the authoritative record of what this dataset contains.

Every image carries provenance: source, source URL, licence, region, capture date, annotator where known, label origin (human / machine / legacy / open-images) and adjudication status. Licences differ per subset and are recorded per image; there is no single licence for the whole repo, which is why the header says other.

Subsets

legacy

frames	370
boxes	403
crops	403
crops pending adjudication	403
bins per frame	{1: 341, 2: 25, 3: 4}
licence	MIT
region	de-by-deggendorf
label origin	legacy

open_images

frames	1110
boxes	1936
crops	0
crops pending adjudication	0
bins per frame	{1: 708, 2: 206, 3: 98, 4: 40, 5: 28, 6: 13, 7: 8, 8: 5, 9: 2, 12: 1, 18: 1}
licence	https://creativecommons.org/licenses/by/2.0/
region	unknown
label origin	open-images

negatives

frames	17474
boxes	0
crops	0
crops pending adjudication	0
bins per frame	{0: 17474}
licence	https://creativecommons.org/licenses/by/2.0/
region	unknown
label origin	open-images

How to read this

Each subset directory is a self-contained pool:

<subset>/
├── manifest.json   provenance for every frame, and every crop
├── images/         full frames
├── labels/         YOLO labels; an EMPTY file means a deliberate background image
└── crops/          identifier candidates (present only where relevant)

Large subsets interpose a shard level – images/ab/<id>.jpg, the two hex characters being sha256(stem) – because the Hub rejects a push where any directory holds more than 10 000 files, and the negative corpus alone is ~17 500. manifest.json says which: "layout": "sharded", and a manifest without the key is flat. legacy/ is flat and stays that way.

sbr.dataset.prepare.build_yolo_tree assembles them into a training tree, grouping frames by capture cluster so that two photographs of the same bin can never straddle a split.

What this data is not

The legacy subset is one city in one week, 370 usable frames, and not one frame of it holds four or more bins — every multi-bin frame here comes from Open Images. Numbers measured on legacy alone are in-distribution numbers.

There is no geographic holdout. Every Open Images frame carries region_id: "unknown", because the source does not record where a photograph was taken, so no split in this dataset answers "does it work in another city". See docs/11-phase2-results.md in the project repo for what was measured, on which split, and on what hardware.