- Jupyter Notebook 100%
| Filename | Latest commit message | Latest commit date |
|---|---|---|
| Figures | ||
| Notebooks | ||
| answers.csv | ||
| LICENSE | ||
| perceptual_scales.csv | ||
| README.md | ||
RAID-Dataset: Human Responses to Affine Image Distortions and Gaussian noise
This repository contains the data, code, and usage examples related to the article:
RAID-Dataset: human responses to affine image distortions and Gaussian noise
Paula Daudén-Oliver*, David Agost-Beltran, Emilio Sansano-Sansano, Raúl Montoliu, Valero Laparra, Jesús Malo, Marina Martínez-García (under review).
🧠 Project Overview
RAID (Responses to Affine Image Distortions) is a perceptual image quality database built from human judgments. Unlike traditional databases focused on digital distortions, RAID investigates suprathreshold affine transformations — rotation, translation, scaling, and Gaussian noise — which are more representative of distortions encountered in natural viewing conditions.
Subjective responses were collected using the psychophysical method Maximum Likelihood Difference Scaling (MLDS). Over 40,000 image comparisons were performed by 210 human observers under controlled laboratory conditions.
📂 Repository Structure
answers.csv: Raw experimental data (individual responses from observers).perceptual_scales.csv: MLDS perceptual scale curves per image, distortion, and distortion level (normalized).images/: Original and distorted images (available via HuggingFace).Notebooks/:Load_DDBB_example.ipynb: Load images and responses.
Load_RAW_data_and_compute_MLDS.ipynb: Compute MLDS curves from raw data.
Load_MLDS_data_and_plot_curves.ipynb: Plot normalized curves.
Convert_MLDS_to_MOS.ipynb: Convert MLDS curves to MOS (aligned with TID2013).
Load_RAW_data_and_plot_left_right_RT.ipynb: Analyze reaction times and decision patterns.
📊 What’s in the Database?
- 888 images: 24 reference images + 864 distorted versions.
- 4 types of distortion:
- Rotation (up to ±18°)
- Translation (up to 0.63° visual angle)
- Scaling (up to ±9%)
- Gaussian noise (matched with TID2013 levels)
- MLDS curves: Quantify perceptual difference as a function of distortion level.
- Reaction times: Provide additional validation via Piéron’s law.
🧪 Technical Validation
Our dataset was validated at multiple levels:
- Reproduction of absolute detection thresholds for each distortion type.
- Compliance with Piéron’s law (reaction time vs. task difficulty).
- Alignment with existing databases (TID2013), and improved performance as shown through Group-MAD experiments.
📥 Data Access
The full dataset and images are available at:
📦 https://huggingface.co/datasets/paudauo/Affine_Transformation_Database
📦 https://doi.org/10.5281/zenodo.15341729
🛠 Requirements
The code has been tested with the following:
- Python 3.10.12
- Numpy 1.26.4
- Pandas 2.2.3
- JAX 0.4.35
📜 License
The dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) License.
🤝 Citation
If you use this dataset or code, please cite the corresponding article (when available). In the meantime, you may reference it as:
Daudén-Oliver, P. et al. RAID-dataset: Human Responses to Affine Image Distortions and Gaussian noise. Scientific Data (under review). GitHub: https://github.com/paudauo/BBDD_Affine_Transformations
🧪 Experimental Design & Results
👁️ Example Trial: MLDS Psychophysical Task
In each trial, participants were shown two pairs of images (quadruple display): one pair on the left and one on the right. Their task was to choose the pair that showed a larger perceptual difference. Each trial involved 4 distorted versions of the same reference image, and trials were designed to balance difficulty and prevent trivial answers.
A total of 210 observers completed over 40,000 trials, under controlled lighting conditions and consistent viewing distance. Reaction times were recorded for every decision, providing insight into perceptual processing.
📈 MLDS Perceptual Scales
The following figure shows the perceptual scale curves for each distortion type (rotation, translation, scaling, and Gaussian noise), computed using the Maximum Likelihood Difference Scaling (MLDS) method.
As distortion level increases:
- Affine distortions (rotation, translation, scaling) show an approximately linear perceptual curves.
- Gaussian noise exhibits a saturating behaviour, aligning with established perceptual models.
