Arbiter-mini

A small, purpose-built image dataset of household items captured under controlled Raspberry Pi camera conditions and labeled for binary waste/recycle classification according to San Diego, CA municipal recycling rules. Built as deployment-condition training data for the Arbiter sorting system, intended to be used alongside TrashNet to close the domain gap between studio imagery and real Pi-camera inference.

Motivation

Models trained purely on TrashNet underperform when deployed on a Raspberry Pi 4 with a Pi Camera Module 3, because TrashNet's capture conditions differ substantially from a fixed overhead Pi camera in a real sorting rig. Arbiter-mini provides a small set of images captured in exactly the conditions a deployed model will see.

Dataset summary

• Total images: 475
• Labels: binary — recycle and trash
• Labeling rules: San Diego, CA municipal recycling guidelines
• Background: white
• Splits: single train split

Capture conditions

• Camera: Raspberry Pi Camera Module 3 (IMX708)
• Mounting: Fixed overhead, ~28–30 cm above the capture surface
• Lighting: LED ring, constant
• Background: white
• Auto white balance: Disabled. AwbEnable: False with hardcoded ColourGains to prevent per-frame color drift that would otherwise create class-correlated confounds.

Locked AWB is the most important capture decision in this dataset. Without it, the camera's white balance shifts subtly per image, and those shifts can correlate with the dominant color of whatever object is in frame — silently leaking class information into color statistics. Anyone reusing this dataset should match these conditions at inference time.

Usage

from datasets import load_dataset

ds = load_dataset("aaryavlal/arbiter-mini", split="train")

To combine with TrashNet, you must first map TrashNet's 6 classes (cardboard, glass, metal, paper, plastic, trash) to binary recycle / trash. The exact mapping used by the Arbiter project (following San Diego rules) is implemented in src/dataset.py.

from datasets import load_dataset, concatenate_datasets

arbiter = load_dataset("aaryavlal/arbiter-mini", split="train")
# trashnet = load and remap to binary using your locality's rules
# combined = concatenate_datasets([arbiter, trashnet])

Relationship to TrashNet

Arbiter-mini is designed to complement TrashNet, not replace or redistribute it. TrashNet files are not contained in this repository. Note that Arbiter-mini uses binary labels (recycle / trash) while TrashNet uses 6 categories — combining the two requires applying a binary mapping to TrashNet, which is locality-specific.

Known limitations

• Binary labels are locality-specific. The recycle / trash mapping reflects San Diego, CA municipal recycling rules. Users in other regions cannot meaningfully relabel without re-examining each image, because the original object category (paper, plastic, etc.) is not preserved in the folder structure.
• Small size. images is enough to shift a TrashNet-trained model toward Pi conditions, not enough to train a strong classifier from scratch.
• Single rig, single environment. One camera, one lighting setup, one background (white), one room.
• Label granularity is coarser than TrashNet. Combining the two requires you to apply your own 6-class → binary mapping to TrashNet before concatenation.
• No human subjects, no PII. Images contain household items only.

Intended use

• Fine-tuning small image classifiers (e.g. MobileNetV3) for binary waste sorting on Raspberry Pi hardware, in combination with a remapped TrashNet.
• Studying domain gap between studio-captured and deployment-captured waste imagery.

Out-of-scope use

• As a general-purpose recycling benchmark — the binary mapping is locality-specific and category-level information is not preserved.
• As a standalone training set — too small and too narrow.
• As a drop-in TrashNet replacement — the label schemas differ.

License

Arbiter-mini is released under CC-BY-4.0. TrashNet is licensed separately (MIT) by Gary Thung and is not included in this repository.

Citation

@misc{arbiter_mini_2026,
  author       = {Aaryav Lal}
  title        = {Arbiter-mini: A Pi-captured companion dataset for binary waste classification},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/aaryavlal/arbiter-mini}}
}

@misc{thung2016trashnet,
  author       = {Gary Thung and Mindy Yang},
  title        = {TrashNet},
  year         = {2016},
  howpublished = {\url{https://github.com/garythung/trashnet}}
}

Acknowledgments

• TrashNet by Gary Thung and Mindy Yang.
• Built as part of the Arbiter project.