ccm/2026-24679-image-dataset

24-679 (Fall 2026): Campus Recycling and Trash Bins

ccm/2026-24679-image-dataset

Classroom photographs of campus recycling and trash bins, prepared as square RGB images with four separately generated training variants. The intended exercise is transfer learning and evaluation on a small image collection, rather than deployment as a waste-sorting system.

Source and task

The preparation notebook uses the Google Forms export Image Data.csv and a ZIP of the form's File responses folder. It matches the counts of links in each upload question with files in that question's folder. 0 = recycling; 1 = trash. Labels are inferred from the upload question/folder, not independently verified from pixels. Count agreement alone does not prove every submitted image was placed under the correct question.

Preparation source: 24-679 Image Data notebook. Course: 24-679, Fall 2026, Carnegie Mellon University. Repository maintainer: the account shown above.

Fields

Field	Meaning and modeling role
image	Prepared RGB pixels; the model's image input.
label	Classification target: 0 = recycling; 1 = trash.
label_name	Readable target description; exclude from model inputs.
source_id, parent_id	Unique prepared/augmented example key and its original source key; provenance only.
augmentation, is_augmented	Method identifier and random-augmentation flag; provenance only.

Splits and original-source counts

These counts are computed from the packaged splits for this run.

Split	Original rows	Synthetic rows	Total rows
train	86	344	430
validation	19	0	19
test	19	0	19

Class counts include synthetic rows in training only.

Split	Label	Rows
train	0	220
train	1	210
validation	0	10
validation	1	9
test	0	10
test	1	9

Requested holdout fraction: 30%; test receives 50% of that holdout. Small-sample rounding changes the realized proportions. The first split uses seed 24679, and the holdout split uses seed 24680.

Split original image indices with class stratification before creating any random variants. Validation and test contain only unaugmented prepared parents. A source and its descendants never cross boundaries. Different photos of the same physical bin, photographer, or location can still cross splits; this is not a grouped-by-bin, participant, or new-location evaluation. Reuse the existing partitions, and do not split the augmented rows again.

Augmentation and preprocessing

Prepared image size in this run: 512 × 512 RGB. Padding color is (128, 128, 128).

All splits receive the same deterministic preparation: apply EXIF orientation, convert to RGB, resize while preserving aspect ratio, and pad to the configured square size. Working copies discard camera metadata; native source files remain separate. Downstream models still need their own checkpoint-specific input processing.

Each original training source contributes its prepared parent and four independent variants; transforms are not stacked. The current notebook uses brightness factors 0.4–0.7 or 1.4–2.0; rotation of 15–30 degrees in either direction on a fixed canvas; contrast factors 0.35–0.65 or 1.5–2.2; and Gaussian blur radius 2.0–4.5 pixels. Lower/higher brightness and contrast ranges are chosen with equal probability. Rotation fills exposed corners with a color derived from the parent and can clip edges.

The stored identifiers mild_brightness, small_rotation, mild_contrast, and mild_gaussian_blur are historical names retained for compatibility. They refer to the stronger settings above, not a measured guarantee of mild distortion. See the linked preparation notebook for the implemented transforms.

Training method	Stored rows
mild_brightness	86
mild_contrast	86
mild_gaussian_blur	86
none	86
small_rotation	86

Intended use and limitations

Use for teaching image preparation, augmentation inspection, and transfer learning. The collection is small and geographically narrow; backgrounds, signage, camera choices, and repeated bins can act as shortcuts. Labels may be mistaken or visually ambiguous. Strong brightness/contrast changes can erase details, rotation can crop evidence, and blur can hide text or recycling symbols. Copied labels and provenance assertions do not prove visual label preservation; inspect parent/variant galleries. Synthetic variants do not add independent photographs. Report class counts, accuracy, and per-class errors alongside weighted F1; use a new-bin or new-location holdout before claiming broader generalization. This dataset does not identify individual waste materials or establish deployment reliability.

Privacy and licensing

The current packaged splits omit original upload filenames. The preparation workflow uses anonymous file IDs, and prepared image bytes have no retained camera metadata. This does not anonymize faces, signage, or other identifying details visible in the photographs, or establish participant consent. Source archives and earlier repository versions may still contain original filenames. Review the images and applicable course permissions before reuse. No license was specified in repository metadata when this card was first added; this card does not assign one.

Load and compare

from datasets import load_dataset
ds = load_dataset("ccm/2026-24679-image-dataset")
# Train with ds["train"], choose settings with ds["validation"], then score ds["test"].

Use an account with access if repository visibility changes. For reproducible comparisons, record the dataset commit and model/environment versions. Regenerate this card with the preparation notebook after changing the data; its counts are calculated from the actual packaged splits. The YAML schema and split configuration are preserved from the upload.