README.md

---
license: cc-by-4.0
task_categories:
  - image-classification
language:
  - en
tags:
  - waste
  - recycling
  - trash
  - garbage
  - environment
pretty_name: Trash Optimizer Dataset
size_categories:
  - 1K<n<10K
---

# Trash Optimizer Dataset

Assembled waste classification dataset covering **18 categories**, used to train the
[cpoisson/trash-optimizer-models](https://huggingface.co/cpoisson/trash-optimizer-models) series.

Built with a custom pipeline that normalises, filters and balances images from three sources:

| Source | Categories contributed | License |
|---|---|---|
| [RealWaste](https://archive.ics.uci.edu/dataset/908/realwaste) (UCI / Kaggle) | cardboard, food_organics, glass, metal, miscellaneous_trash, paper, plastic, textile_trash, vegetation | CC BY 4.0 |
| [RHWC](https://www.kaggle.com/datasets/alistairking/recyclable-and-household-waste-classification) (Kaggle) | cardboard, food_organics, glass, metal, paper, plastic, textile_trash | CC BY-SA 4.0 |
| Custom (web-scraped) | battery, car_battery, light_bulb, mirror, neon, pharmacy, printer_cartridge, tire, wood | ⚠️ Mixed — images sourced from the public web, licenses vary |

> **Note**: The custom category images were collected from publicly accessible web sources.
> Their individual licenses may vary. Use for research and non-commercial purposes.

## Structure

```
dataset/
  battery/          277 images
  car_battery/      260 images
  cardboard/        500 images
  food_organics/    500 images
  glass/            500 images
  light_bulb/       157 images
  metal/            500 images
  mirror/           194 images
  miscellaneous_trash/ 500 images
  neon/             190 images
  paper/            500 images
  pharmacy/         257 images
  plastic/          500 images
  printer_cartridge/ 357 images
  textile_trash/    500 images
  tire/             353 images
  vegetation/       436 images
  wood/             245 images
```

**Total: 6,229 images across 18 categories**

Images are JPEG/PNG, minimum size 150×150px, maximum 500 images per category.
Filename prefix indicates source: `realwaste_*`, `rhwc_*`, `customdataset_*`.

## Dataset builder

The assembly pipeline is available in the
[trash-optimizer](https://github.com/cpoisson/trash-optimizer) repository
under `dataset/datasetbuilder.py`, configured via `trash-optimizer-datasetbuilder.toml`.

## Model trained on this dataset

→ [cpoisson/trash-optimizer-models](https://huggingface.co/cpoisson/trash-optimizer-models)
— EfficientNet-B0, 95.07% test accuracy across 18 categories.

## Citation

If you use the RealWaste portion, please cite:
```bibtex
@misc{realwaste,
  author    = {Majchrowska, Sylwia and others},
  title     = {RealWaste},
  year      = {2023},
  publisher = {UCI Machine Learning Repository},
  doi       = {10.24432/C5SS4S}
}
```