Waste classification image dataset COCO: the annotated corpus and the shelf behind it

A waste-sorting or litter-detection model needs two things from its training material: photographs of trash where trash actually lives, and per-object labels in a schema the pipeline already reads. The COCO format supplies the second half - JSON that stores, for every image, a list of objects with category IDs and polygon masks instead of one class label for the whole frame.

That distinction matters more in waste work than in most vision niches. A street photograph routinely holds several litter items of different materials at once, so whole-image classification throws away exactly the information a sorting line or pickup-routing application needs. Instance segmentation - polygon-level labels separating each item - is what lets a model count, locate and classify individual pieces of trash in one pass.

Two records in Datadory's sixteen-record Environmental & Facilities Services shelf do this work directly: TACO - Trash Annotations in Context (scored 7/10), the curated corpus with native COCO-format polygons, and Hugging Face Waste Datasets (scored 7/10), the 273-repository community shelf spanning imagefolder, webdataset and tabular formats. Everything else in the slice measures facilities, emissions or statistics - useful as ground truth for what a deployed model claims, useless as training pixels.

TACO is the only curated image record on this shelf with native COCO-format instance segmentation. Its published numbers are small but precise: 1,500 annotated photographs, 4,784 segmentation annotations and a 60-category hierarchical waste taxonomy, packaged into an archive of roughly 2.7 GB. Every photograph shows litter where people encounter it - woods, roads and beaches - rather than on a studio conveyor belt.

One row explains why the format matters more than the headline count:

image_id=0   file=batch_1/000006.jpg   size=1537x2049
category=Clear plastic bottle
bbox=[517.0, 127.0, 447.0, 1322.0]
segmentation=[[561.0, 1238.0, 568.0, 1201.0, ...]]
area=403954 px

The mask covers about 12.8 percent of a 3.15-megapixel frame - the bottle dominates its scene rather than hiding in it. The polygon gives an exact outline, not a loose rectangle, and category_id resolves into the 60-class hierarchy while scene_annotations keeps background context beside the objects themselves. Across the corpus those records total 1,500 images, 4,784 segmentations and 60 categories, frozen as version 1.0 in December 2019.

Three structural facts shape how far it stretches. Class balance is uneven by nature: cigarette butts and bottles have plenty of examples while most fine-grained classes have few, so published class maps merge rare categories into detector-friendly groupings such as cans, bottles and plastic bags. No geographic metadata rides with individual images, so country-level claims need a companion source. And the reviewed corpus froze at v1.0 while community annotations accumulate weekly without curator review - reproducible benchmarking anchors on v1.0 and treats anything newer as provisional. Datadory ships both layers separately and clearly flagged.

The taxonomy runs from whole objects down to their components, which is why 4,784 annotations exceed what a flat 1,500-image set suggests. A single photographed beverage can contribute several labelled instances - the can body, its pulled tab, its plastic wrapper - each earning its own polygon and category ID in the COCO file.

Treat that hierarchy as a design decision you inherit. Training a 60-way classifier spreads 4,784 instances thinly; collapsing sibling categories into material-level superclasses trades label resolution for per-class sample depth. Because the labels are polygons keyed by category ID rather than fixed bounding boxes, either strategy falls out of the same annotation file with no re-labelling work - deciding between full resolution and merged grouping is the first scoping conversation when you request a sample.

273 publicly listed repositories answer 'waste' on the Hugging Face shelf as of August 2026. It is not one collection but a community index, dominated by computer-vision material built for sorting and recycling automation: TACO-derived waste-recognition sets, organic-versus-anorganic classification folders, detect-waste object-detection corpora, GreenLabel waste-type datasets and drone imagery of waste dumpsites. Five entries show the range:

Quality is the honest deduction on this shelf's 7/10 score: anything public can be uploaded, so provenance runs from disciplined organisational releases to single-author hobby projects, and several repositories turn out to be mirrors of TACO itself. Uploads span roughly 2021 to the present, so the shelf reflects whatever is currently public rather than a curated snapshot. Datadory vets each repository individually before it reaches a delivery and preserves uploader attribution on every row. For the trade-offs side by side, see the head-to-head comparison.

Volume context arrives beside them. Eurostat Waste statistics deliver harmonised EU generation and treatment series - municipal records reaching back to 1995, 2024 figures released March 2026 - and the OECD municipal waste statistics cover 58 reference areas for cross-country panels.

Six steps take a team from empty directory to a defensible first detector:

1. Start from the frozen benchmark. Train and evaluate against TACO's 4,784 reviewed masks over the fixed 60-class taxonomy - a frozen test set is the only honest way to compare two models' numbers.

1. Decide the label collapse before training. Map the hierarchy onto whatever class set the application needs: full resolution for research, material-level superclasses when 4,784 instances spread thin. Both derive from the same polygon annotations with no re-labelling.

1. Split by capture context, not randomly. Near-duplicate frames leak between train and validation splits; hold out whole scenes so the metric reflects generalisation to unseen environments.

1. Augment from the shelf selectively. Pull the dumpsite drone imagery, food-waste detection sets and TACO-derived recognition folders whose reuse terms read clean at scoping, and skip anything unvetted unless the legal exposure is yours alone to carry.

1. Wire in the ground-truth layer. When predictions drive user-facing advice, join them against ECHO's 1.5-million-plus facility rows or Defra's carriers register, and cite Eurostat or OECD series for any volume claim.

Litter-detection model builders pretrain and benchmark on the reviewed masks, then widen coverage from the community shelf once the baseline survives contact with real clutter.

Smart-bin, sorting and environmental-product teams prove feasibility before hardware commitments: if a detector cannot separate cigarette butts from bottle caps on this corpus, the production environment will not be kinder.

Municipal and NGO analysts quantify what actually litters different environments - the 60-class breakdown turns 'there is a lot of plastic' into countable categories, and the tonnage series put counts beside mass.

Competitive-intelligence teams scope what training material already exists before budgeting bespoke annotation - drone dumpsite detection included - which is a sourcing question the shelf answers in an afternoon.

The personas who lean hardest: data scientists and developers and builders lead, with journalists and academics close behind for citable, documented corpora.

Because the hard part was never knowing these corpora exist - it is reconciling them. Two annotation conventions that only partly overlap, a taxonomy designed in 2019 meeting community uploads from last week, reuse declarations that change per uploader, nulls that mean absence rather than zero, and a frozen benchmark sitting beside a moving index. Each quirk is documented somewhere; together they eat a week.

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

Pick the channel your stack already speaks; the pixels travel unchanged through all three with annotation structure intact, and the statistical sides land as typed rows against documented dictionaries - so tonnage series never mix with mask counts and missing readings arrive flagged rather than interpolated.

For teams assembling this stack by hand, the honest accounting is weeks: a frozen benchmark, a 273-item vetting queue and four statistical publishers whose calendars share nothing. That assembly work is ours, which leaves the interesting work - the model, the sorter, the deployment - to you.

Start with the environmental facilities services data guide, the pillar post for this industry: it ranks all 19 pooled records - 16 primary plus 3 related - with quality scores and workflow groupings, placing this computer-vision cluster beside the compliance, benchmarking and statistics clusters.

Product-level detail - sample rows, field dictionaries, coverage chips - lives on the pages behind this piece: TACO - Trash Annotations in Context and Hugging Face Waste Datasets. The data scientists use cases page maps which records suit modelling versus retrieval work, the environmental facilities services industry hub lists every cataloged record in one place, and the best environmental-facilities-services datasets ranking shows what to automate first.

• environmental & facilities services datasets
• free environmental-facilities-services datasets
• best environmental-facilities-services datasets
• Hugging Face vs TACO comparison