Description
ONOT is a synthetic dataset specifically focused on the generation of high-quality faces in adherence to the requirements of the ISO/IEC 39794-5 standards. Following the guidelines of the International Civil Aviation Organization (ICAO), this ISO/IEC standard defines the interchange formats of face images in electronic Machine-Readable Travel Documents (eMRTD).
We take this standard as the inspiring principle for our generation process, which is aimed at the creation of high-quality and well-controlled images with specific characteristics including, among others, frontal face pose with uniform background and illumination, neutral expression, and the absence of shadows.
Realization
ONOT dataset has been created following a four-step procedure:
- Image Generation: the generation is based on a fine-tuned version of Stable Diffusion 1.5, namely Realistic Vision 5.1. The model is served using Stable Diffusion Web UI. Each image has a resolution of 512 × 512 and is generated using the DPM++ SDE Karras sampler with 25 steps.
- ISO/IEC 39794-5 compliance: the compliance verification procedure is carried out through a commercial SDK. Specifically, this SDK verifies
the presence of scene constraints (such as pose, and expression), photographic properties (e.g. lighting, positioning, and camera focus), as well as digital image attributes (e.g. image resolution, and image size). - Intra-class consistency: here, we aim to verify if the images grouped in each pseudo-class belong or not to the same identity.
- Inter-class consistency: as the prompts for the different pseudo-classes may generate subjects that are too similar to each other, the next step is to select the pseudo-classes that contain faces that are all dissimilar enough.
ICAO tests
Main Features:
- Different genders, ethnicities, ages and facial traits
- European / American (EEA)
- African (EAF)
- Indian / Asian (EIA)
- East-Asian (EAS)
- Middle Eastern (EME)
- High variety in appearances, hairstyles, and accessories.
Release
The inter- and intra-class consistency tests are based on three thresholds for the face verification systems, and therefore producing three different subsets of data. These distinct subsets correspond to different working scenarios.
- Subset 1: low threshold (FMR 100 = 0.413)
- Low intra-class variability, and then fewer images for each subject
- High level of inter-class similarity, less variability, and then more number of subjects
- 255 identities, 5707 images
- Scenario: a challenging benchmark for face analysis tasks since there are a large number of look-alike subjects
- Subset 2: medium threshold (FMR 1000 = 0.493)
- Medium level of similarity, medium intra-class variability
- 125 identities, 4729 images
- Subset 3: high threshold (FMR = 0.597)
- High level of intra-class variability, and then more images for each subject
- Low level of inter-class similarity, more variability (distinct subjects), and then fewer number of subjects
- 55 identities, 2853 images in total
- Scenario: improving the robustness of FRSs to typical variations of face appearance
Data
In the download folder, there are three .txt files that contain the names of the images belonging to each subset.
For each line, in the first position, there is the name of the ICAO-compliant image.
To download the dataset, click the button!
For further information, please send an email to one of the authors.
License and Reference
The ONOT dataset is released under the CC BY-NC 4.0 license: you may not use the material for commercial purposes and you must give appropriate credit (see the reference below). You are free to download and modify the dataset.
If you use this dataset please cite the following paper:
@inproceedings{di2024onot,
title={ONOT: a High-Quality ICAO-compliant Synthetic Mugshot Dataset},
author={Di Domenico, Nicol{`o} and Borghi, Guido and Franco, Annalisa and Maltoni, Davide and others},
booktitle={The 18th IEEE International Conference on Automatic Face and Gesture Recognition (FG)},
pages={1--6},
year={2024}
}