KPatelis/oxford-iiit-pet-clean

Oxford-IIIT-Pet Cleaned

This dataset is a cleaned and enriched version of the timm/oxford-iiit-pet dataset.
Ideally suited for fine-grained image classification tasks, this version addresses common dataset quality issues such as duplicates, corrupt files, and leakage between train/test splits, while preserving the original data structure where appropriate. Additionally, the train set has been further split into train_clean and validation_clean datasets using an 80/20 split ratio stratified across pet breeds.

Dataset Description

The original Oxford-IIIT Pet Dataset is a 37-category pet dataset with roughly 200 images for each class. The images have large variations in scale, pose, and lighting.

This Cleaned Version offers:

• Deduplication: Removal of exact and near-duplicates within splits.
• Leakage Prevention: Removal of training samples that are near-duplicates of testing samples.
• Quality Annotation: A new issues column that flags potential problems (blur, low contrast, label consistency, etc.) without removing the data, allowing users to filter based on their own criteria.
• Health Checks: Removal of corrupt or unreadable images.

Supported Tasks

• Fine-Grained Image Classification: Distinguishing between 37 breeds of cats and dogs.
• Data Quality Research: analyzing the impact of dataset cleaning on model performance.

Dataset Structure

The dataset contains four splits:

Split	Description
train_clean	80% of the training set with all critical issues removed (Corrupt, Internal Duplicates, Cross-Split Leakage). Label errors and quality issues are kept but annotated in the issues column.
validation_clean	20% of the training set with all critical issues removed (Corrupt, Internal Duplicates, Cross-Split Leakage). Label errors and quality issues are kept but annotated in the issues column.
test_clean	The test set with internal duplicates and corrupt files removed. Cross-split duplicates are handled by removing the corresponding image from train_clean to preserve value of the standard test benchmark.
test_original	The original test set, unaltered in terms of rows, but enriched with the issues column for analysis. Allows for comparison against other models finetuned on this task.

Data Fields

• image: A PIL.Image.Image object containing the image.
• label: An int classification label (0-36).
• image_id: A unique identifier for the image.
• label_cat_dog: An int classification label (0 or 1), either cat or dog.
• issues: A list of strings detailing issues found in the image.

The issues Column

The issues column is generated by an automated cleaning pipeline and may contain the following tags:

• Corrupt file: The image file could not be decoded (these are removed from clean splits).
• Health: <issue>: Structural issues like extreme aspect ratios or very small dimensions.
• Quality: <issue>: Visual quality issues such as Blurry, Low Contrast, Dark, or Bright.
• Label Error: Semantic Inconsistency: The image embedding is significantly different from others in its class (potential mislabel or outlier).
• Internal Duplicate of <id>: The image is a duplicate of another image in the same split.
• Leakage (Duplicate of Test <id>): The image is a duplicate of an image in the test split.

Citation

If you use this dataset, please cite the original authors:

@InProceedings{parkhi12a,
  author       = "Omkar M. Parkhi and Andrea Vedaldi and Andrew Zisserman and C. V. Jawahar",
  title        = "Cats and Dogs",
  booktitle    = "IEEE Conference on Computer Vision and Pattern Recognition",
  year         = "2012",
}

Credits

• Original Dataset: timm/oxford-iiit-pet
• Cleaning Script: TODO

License & Disclaimer

This dataset is distributed under the same license as the original Oxford-IIIT Pet Dataset (CC-BY-SA 4.0). We provide no warranty on the dataset, and the user takes full responsibility for usage.