Household Paper Products Occlusion Image Dataset

It is a vision corpus aimed at the least glamorous failure in retail AI: the pack of bath tissue photographed behind the shelf edge. Publisher Mobiusi states the problem plainly on the dataset card - retail e-commerce struggles to detect household paper products because of packaging variety and occlusion - and answers it with high-resolution photography captured in controlled environments, annotated across multiple rounds with expert review. The card builds the corpus around Kleenex, Scott, Charmin and Tempo, and claims tissue-and-towel breadth alongside the toilet-paper core.

The committed slice is modest, and reading it correctly is most of the work. The repository, researched during the August 2026 pass, was created on 2026-03-05 and last touched the same day, and commits four JPG images with JSON annotation sidecars - about 1.31 MB - in a single train split under one default subset. Downloads last month: 18. Read together, those facts describe a fully specified labeling grammar shipped with a small working demonstration, and a fuller run whose scope sits with the publisher. Inside the catalog this record anchors the paper half of the household-products data hub computer-vision column.

One of the four committed annotation records, reproduced verbatim - which means this block displays a quarter of the entire committed corpus, itself the most useful fact on the page:

# committed annotation record, shown exactly as it lands

file_name       : 3a8bdb8b15cad1d2c994646ef395ed23.jpg
quality         : 1280*1431
object_type     : toilet paper
occlusion_level : unobstructed

# published label vocabularies (from the auto-converted viewer, not extra rows)

object_type     : toilet paper                          # value on all four committed rows
occlusion_level : unobstructed | partially occluded

Read it as the control case of a designed pair of extremes. A portrait-framed 1280*1431 capture, product fully visible, labeled unobstructed - while the vocabulary above it names the opposite pole, partially occluded, that a detection model actually needs to see to learn. And note what the object_type vocabulary admits: all four committed rows read toilet paper, so the tissue-and-towel breadth claimed on the card is an assertion about the fuller corpus, not something the public slice evidences. Both observations belong in any scoping conversation.

Four fields make up the verified dictionary, every definition checked against the committed annotation records during the August 2026 research pass - no invented columns padding the table. They split into two jobs:

• Identity and quality: file_name addresses each asset with a hash-style identifier tying the JPG to its JSON sidecar, while quality records capture resolution as width*height - the committed row observes 1280*1431, a portrait crop worth knowing about if your input pipeline resizes aggressively.
• Labels: object_type names the product pictured and occlusion_level grades how much of it shows. These two carry the entire training signal, which makes their consistency the thing to manage.

That consistency note deserves its own sentence: the card's technical spec table describes occlusion_level as a float between zero and one, while the viewer publishes categorical values (unobstructed, partially occluded). Both descriptions belong to the same attribute. Anyone joining this corpus to its cleaning-products sibling - which uses title-case categories such as Partially Occluded - should normalize the convention first, because a naive string match fails silently across the pair. The full dictionary follows below; extended assets ride along with any scoped sample on request.

• Geography - none, stated as a feature: these are controlled-environment studio photographs rather than store surveys, so there is no regional weighting, no retailer mix and no store-format bias baked into the pixels. Occlusion is a global retail problem; the corpus treats it that way.
• Temporal - a fixed snapshot. Created 2026-03-05, last modified the same day, classified static. Nothing revises underneath a trained model, and equally, new packaging generations do not arrive on their own.
• Granularity - one annotation record per image, four labeled attributes each, in a single train split with no published validation or test partitions. Stratification is left to the consumer by design.
• Scale - the gap to plan around. The card describes a multi-brand paper-products corpus with expert review; the committed slice holds four images totaling about 1.31 MB. Both numbers are reported because capacity planning needs both, and intended corpus scope is confirmed explicitly at sample request.

Against the wider catalog - a mean quality score of 7.81 across 1,744 datasets - this record scores 4/10, a band shared by only 30 records, and its static cadence matches 181 others. Where it ranks on the shelf: best household products datasets.

API, files, or your warehouse. Daily, weekly, or hourly.

You choose the channel and the cadence; the labeling grammar travels intact either way. Images arrive matched to their JSON sidecars by filename stem, or flattened into tabular rows if your pipeline prefers DataFrames to file trees, and the four-field dictionary stays byte-stable across deliveries - a loader written against the sample above runs unchanged against whatever slice you scope.

Because the upstream corpus is itself frozen, cadence here governs how current your copy of the delivery pipeline is, not how fresh the photographs are - worth deciding deliberately rather than defaulting to hourly. For model-development sprints a one-time extract lands once and never moves. Either way the sample ships first: real rows in your named categories before any commitment.

Five jobs this record settles outright:

1. Paper-aisle shelf-recognition pilots - paired object_type and occlusion_level labels give a detector graded difficulty instead of one easy case; see ML model training use cases.
2. Catalog imagery QA for e-commerce - classifiers exposed to partially-occluded examples flag supplier photos where wrapping, shelf edges or neighboring packs hide the pack before it reaches a live listing.
3. Annotation-schema prototyping - four crisply defined fields are a proven minimal template for teams standing up their own capture-and-label operation for household SKUs, whether they keep it or argue with it.
4. Robustness benchmarking - score candidate models separately on the unobstructed and partially-occluded strata and report the delta honestly; that number survives contact with a real aisle.
5. Whole-basket vision coverage - joined with its cleaning-products sibling, the pair covers the two most occlusion-prone aisles of the household basket under near-identical grammars.

Each job maps to a persona below, and the sample validates whichever one you came for.

Data scientists and ML engineers take the highest fit: purpose-built imagery with a documented four-field schema compresses the distance from question to trained detector; see data scientists household products use cases. E-commerce operators inherit the card's stated problem directly - listing imagery that fails recognition costs conversion. Developers and builders get hash-named files with JSON sidecars that drop into prototypes without glue code; see developers and builders personas. Competitive intel and product teams should calibrate expectations: this is a vision-training layer with no sales velocity, no pricing and no retailer geography anywhere in it; see competitive intel product teams use cases.

Fit has an edge worth naming: pixels, not panels. Commercial questions about the paper aisle - price drift, certification status, ingredient detail - belong to the tabular records beside it on the shelf below.

Three notes worth having upfront.

First, scope the corpus explicitly. The distance between the card's brand roster and the four committed rows is the defining fact of this record, so say which you expect: the demonstrated labeling grammar, the committed slice, or the fuller run - and treat the answer as part of the negotiation rather than an assumption. If tissue and towel coverage matters to you, name it; the public evidence covers toilet paper only.

Second, pick an occlusion convention and keep it. This corpus publishes categorical levels while its own card spec describes a zero-to-one float, and its sibling uses differently-cased category names. Deliveries can normalize to whichever convention you name - but only if you name it.

Third, plan your own splits. With a single train partition published, any honest evaluation stratifies on occlusion_level yourself - trivial with the dictionary above, painful to retrofit after a model has seen everything. Name your categories, occlusion levels and target format, and the sample returns in exactly the schema shown above. Then Get a sample of this dataset.

On the household-products data hub shelf this record owns the paper half of the computer-vision column. Its nearest neighbor is the Household Cleaning Products Occlusion Image Dataset - same publisher, same controlled-environment approach, applied to Lysol, Clorox, Mr. Clean and Vanish instead of Kleenex, Scott, Charmin and Tempo; run together they cover both occlusion-prone aisles. Around them: the Consumer Product Information Database carries ingredient and usage detail at product level, the EPA Safer Choice Certified Products Database lists which formulations earn the safer-choice mark, and the BLS Price & Inflation Data Tools record prices the aisle monthly back to 1997. Seven records total, one shelf - request samples in the same schema and they land in one warehouse.