zacCMU/2025-24679-image-dataset

🏞️ Nature vs. Not Nature Image Dataset

📝 Dataset Summary

This is a student-created image dataset designed for binary image classification. The dataset consists of 31 original photographs, each manually labeled as either depicting a nature scene or not_nature (e.g., man-made objects, indoor scenes).

To facilitate model training, a larger augmented split is provided, expanding the dataset to 310 images through a series of documented, label-preserving transformations. The goal of the task is to train a model that can accurately classify a given image into one of the two categories.

---

Supported Tasks

• image-classification: The primary intended task is to train a model to distinguish between "nature" and "not nature" images. The image column contains the input image, and the label column is the target for prediction.

---

Dataset Structure

Data Instances

The dataset contains images and their corresponding labels. A sample contact sheet of the original 31 images is included in the repository for quick visualization.

Data Fields

• image: A PIL Image object containing the photograph.
• label: A ClassLabel with two possible values: 0 (not_nature) or 1 (nature).

Data Splits

The dataset contains two splits:

• original: Contains the 31 original, manually photographed and labeled images. This split is ideal for validation, testing, or quick prototyping.
• augmented: A larger training set of 310 images. This includes the 31 original images plus 279 new images created via augmentation, providing a more substantial dataset for training a model.

---

Dataset Creation

Curation Rationale

This dataset was created as part of a student project to provide a hands-on example of a complete machine learning workflow, including data collection, manual labeling, data augmentation, and model training. The topic was chosen for its clear binary distinction and ease of collecting sample images.

Source Data Collection and Preprocessing

The 31 images in the original split were personally photographed by the student creator in and around Pittsburgh, Pennsylvania. The collection includes a mix of outdoor nature scenes (parks, trails, foliage) and non-nature scenes (room interiors, everyday objects, buildings).

Before being included in the dataset, each original image was preprocessed with a center-crop to achieve a square aspect ratio, and then resized to a uniform dimension of 224x224 pixels.

Augmentation

The augmented split was generated programmatically to expand the small initial dataset. For each of the 31 original images, 9 new, unique variants were created using a pipeline of label-preserving transformations from the torchvision library. This process helps expose the model to a wider variety of visual data and reduce overfitting. The augmentation techniques included:

• RandomResizedCrop
• RandomHorizontalFlip
• ColorJitter (adjusting brightness, contrast, and saturation)
• RandomRotation

Labels

Each of the 31 original images was manually assigned a binary label (0 for not_nature, 1 for nature) by the dataset creator. The labels for the augmented images were inherited directly from their original source image.

---

Considerations for Using the Data

Intended Use

This dataset is intended for educational and experimental purposes. It serves as a practical, small-scale benchmark for learning and practicing image classification techniques, from data preparation to training a convolutional neural network (CNN).

Limitations and Biases

• Geographic Bias: All photographs were taken in a single geographic region (Pittsburgh, PA) and may not represent nature or non-nature scenes from other parts of the world.
• Photographer Bias: The dataset reflects the artistic style, subject choices, and technical skill of a single student photographer.
• Small Scale: The foundational dataset of 31 unique images is very small, and models trained on it may not generalize well to a wider variety of real-world images without the augmentation.

Ethical Considerations

The dataset contains no personally identifiable information. The images are of general scenes and objects and their use raises no known ethical concerns.

AI Usage Disclosure

The augmented split of this dataset was created using a programmatic, non-generative augmentation pipeline. No generative AI models were used in the creation of the images or labels.

---

Additional Information

Licensing

The dataset is licensed under the MIT License.

Citation

@dataset{nature_vs_not_nature_2025,
  author    = {Zachary Zdobinski},
  title     = {Nature vs. Not Nature},
  year      = {2025},
  url       = {[https://huggingface.co/datasets/zacCMU/2025-24679-image-dataset](https://huggingface.co/datasets/zacCMU/2025-24679-image-dataset)}
}