trojblue/AVA-aesthetics-10pct-min50-10bins

AVA Aesthetics 10% Subset (min50, 10 bins)

This dataset is a curated 10% subset of the AVA Aesthetics Dataset (or the original AVA dataset as described in Murray et al., 2012). It includes images that have at least 50 total votes and have been stratified into 10 bins based on their computed mean aesthetic scores.

Dataset Overview

• Dataset Name: AVA Aesthetics 10% Subset (min50, 10 bins)
• Subset Size: 10% of the original AVA dataset (after filtering for a minimum of 50 votes per image)
• Filtering Criteria: Only images with ≥50 votes were considered.
• Stratification: Images were binned into 10 equally spaced intervals across the score range (1 to 10), and a random 10% sample was selected from each bin.
• Image Format: JPEG (files with .jpg extension)
• Data Fields:
  • image_id: Unique identifier of the image.
  • image: The image file (loaded as an Image feature in Hugging Face Datasets).
  • mean_score: The mean aesthetic score computed from the rating counts.
  • total_votes: Total number of votes received by the image.
  • rating_counts: A list representing the count of votes for scores 1 through 10.

Dataset Creation Process

1. Parsing the AVA.txt File:

   • Each line in AVA.txt contains metadata for an image including the image ID and counts for ratings 1 through 10.
   • Total votes and the mean score are computed as:
     • Total Votes: Sum of counts for ratings 1–10.
     • Mean Score: ( \text{mean_score} = \frac{\sum_{i=1}^{10} i \times \text{count}_i}{\text{total_votes}} ).

2. Filtering:

   • Images with fewer than 50 total votes are removed to ensure label stability and reduce noise.

3. Stratification:

   • The images are grouped into 10 bins based on their mean score. This stratification helps maintain a balanced representation of aesthetic quality across the dataset.

4. Sampling:

   • From each bin, 10% of the images are randomly selected to form the final subset.

5. Conversion to Hugging Face Dataset:

   • The resulting data is transformed into a Hugging Face Dataset with the following fields: image_id, image, mean_score, total_votes, and rating_counts.
   • The dataset is then split into train and test sets (default split: 90% train, 10% test).

Intended Use Cases

• Aesthetic Quality Assessment: Developing models to predict image aesthetics.
• Computer Vision Research: Studying features associated with aesthetic judgments.
• Benchmarking: Serving as a balanced subset for rapid experimentation and validation of aesthetic scoring models.

Limitations and Considerations

• Subset Size: Being only 10% of the original dataset, this subset may not capture the full diversity of the AVA dataset.
• Sampling Bias: The stratification and random sampling approach might introduce bias if certain bins are underrepresented.
• Missing Files: Some images may be absent if the corresponding JPEG file was not found in the provided directory.
• Generalization: Models trained on this subset should be evaluated on larger datasets or additional benchmarks to ensure generalization.

How to Use

You can load the dataset directly using the Hugging Face datasets library:

from datasets import load_dataset

dataset = load_dataset("trojblue/ava-aesthetics-10pct-min50-10bins")
print(dataset)

Citation

If you use this dataset in your research, please consider citing the original work:

@inproceedings{murray2012ava,
  title={AVA: A Large-Scale Database for Aesthetic Visual Analysis},
  author={Murray, N and Marchesotti, L and Perronnin, F},
  booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
  pages={3--18},
  year={2012}
}

License

Please refer to the license of the original AVA dataset and ensure that you adhere to its terms when using this subset.