metadata

dataset_info:
  features:
    - name: image_id
      dtype: string
    - name: image
      struct:
        - name: bytes
          dtype: binary
        - name: path
          dtype: string
    - name: mean_score
      dtype: float32
    - name: label
      dtype: int64
    - name: total_votes
      dtype: int32
    - name: rating_counts
      sequence: int32
    - name: edge_density
      dtype: float64
    - name: focus_measure
      dtype: float64
    - name: texture_score
      dtype: float64
    - name: noise_level
      dtype: float64
    - name: saturation
      dtype: float64
    - name: contrast
      dtype: float64
    - name: brightness
      dtype: float64
    - name: avg_dynamic_range
      dtype: float64
  splits:
    - name: train
      num_bytes: 2737038380
      num_examples: 20437
  download_size: 2710920619
  dataset_size: 2737038380
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

AVA Subset with Metrics

This dataset is a processed subset of the AVA (Aesthetic Visual Analysis) dataset, derived from trojblue/AVA-aesthetics-10pct-min50-10bins. It includes a selection of images alongside computed visual quality metrics.

Derivation Process

Subset Selection: Images were extracted from trojblue/AVA-aesthetics-10pct-min50-10bins, ensuring a minimum of 50 samples per bin.
Efficient Local Export: Images were stored locally using a multi-threaded approach to speed up processing.
Metric Calculation: Various computer vision metrics were computed using cv2_metrics from procslib, including sharpness, contrast, and other image quality indicators.
Data Merging: The computed metrics were merged back into the dataset, providing additional insights beyond aesthetic scores.

Usage

This dataset is ideal for:

Training models that incorporate both aesthetic scores and image quality metrics.
Analyzing relationships between image structure and subjective ratings.
Benchmarking computer vision models on real-world aesthetic quality assessment.

The dataset is publicly available for research and model development. 🚀