Dataset Card for ccm/2025-24679-image-dataset

Dataset Details

Dataset Description

This dataset consists of images labeled as recycling (0) or trash (1). It was created as part of a classroom exercise in supervised learning and data augmentation, with the goal of giving students practice in building and evaluating image classification pipelines.

• Curated by: Fall 2025 24-679 course at Carnegie Mellon University
• Shared by [optional]: Christopher McComb
• License: MIT
• Language(s): N/A (image dataset)

Uses

Direct Use

• Training and evaluating image classification models (binary classification: recycling vs. trash).
• Experimenting with image preprocessing (resizing, normalization, augmentation).
• Teaching end-to-end ML workflows: data loading, training, validation, and evaluation.

Out-of-Scope Use

• Production deployment in real recycling or waste-sorting systems.
• Generalization to real-world trash/recycling classification without larger and more diverse datasets.
• Use in safety-critical or automated decision-making contexts.

Dataset Structure

The dataset includes two splits:

• original: 174 examples (collected by students).
• augmented: 522 examples (synthetically generated to balance and expand the dataset).

Each row includes:

• image: an image file (e.g., JPEG/PNG).
• label: integer class label (0 = recycling, 1 = trash).

Dataset Creation

Curation Rationale

The dataset was curated to provide a simple, hands-on dataset for practicing image classification methods in an educational setting. Recycling/trash was chosen because it is easy to photograph and conceptually straightforward.

Source Data

Data Collection and Processing

• Original images were collected on campus by students (e.g., photographs of bins, bottles, cans, paper, etc.).
• Labels were assigned manually during the collection process.
• Augmented data was generated with transformations such as rotations, flips, brightness/contrast changes, and cropping.

Who are the source data producers?

• Original data: Students in the 24-679 course.
• Augmented data: Generated by course instructors and teaching assistants using standard augmentation tools.

Bias, Risks, and Limitations

• Small sample size: Only 174 original images.
• Synthetic augmentation: Does not capture real-world variation in lighting, backgrounds, or object diversity.
• Domain bias: Limited to CMU campus items, not representative of recycling/trash globally.

Recommendations

• Use primarily for teaching and demonstration purposes.
• Do not generalize beyond the dataset scope.
• Highlight dataset limitations during instruction to reinforce lessons about data quality and bias.

Dataset Card Contact

Christopher McComb (Carnegie Mellon University) — [email protected]