Skip to content

VAMONOT-S: a fully synthetic benchmark for Video-based Morphing Attack Detection (V-MAD)

VAMONOT-S is the synthetic counterpart of VAMONOT-R and has been developed to provide a controlled, privacy-preserving benchmark for evaluating biometric systems under realistic verification conditions.
This dataset has been introduced for the Every Frame You Take: The Video-based Morphing Attack Detection Challenge.

Description

Built from photorealistic synthetic identities, it pairs ICAO-compliant document images with AI-generated (Hailuo AI) video sequences simulating real identity verification scenarios. By combining privacy-friendly data with controlled variations in pose, motion, and acquisition conditions, VAMONOT-S provides a reliable benchmark for training and evaluating next-generation morphing attack detection methods.

The videos are designed to simulate realistic identity verification scenarios, including:

  • Controlled verification scenarios, similar to an airport security checkpoint, where the subject walks toward a fixed camera while maintaining a frontal view.
  • Less constrained scenarios, characterized by natural head and body movements.

VAMONOT-S is built from 55 synthetic identities sourced from the ONOT synthetic dataset of ICAO-compliant facial images.

Task

  • Video-based Morphing Attack Detection (V-MAD)
  • Single-image Morphing Attack Detection (S-MAD)
  • Differential Morphing Attack Detection (D-MAD)

Data Organization

VAMONOT-S follows the same core structure as VAMONOT-R, based on the comparison between a document image and a live video sequence.

The content of the dataset is organized as follows:

  • The whole dataset is divided into 55 folders, one folder for each subject
  • Each subject folder contains:
    • icao: the folder with the ISO/ICAO compliant image
    • morphed: the folder with morphed images
    • videos: the folder with videos
  • Each video folder contains:
    • video_gate: controlled scenario, frontal position of the subject
    • video_gate_cappello_occhiali: less controlled scenario, subject with garments
    • video_in_the_wild: more challenging scenario, with variable head poses, camera positions and garments

Each video folder has the following structure:

  • face_detection: detected faces (computed using MediaPipe)
  • frames: original frames of the video (a frame was sampled every 2)
  • landmarks: annotations and images with landmarks (computed using MediaPipe)
  • ofiq: face quality score computed using the OFIQ framework
Download VAMONOT-S

Citation

When using VAMONOT-S in your research, please cite the corresponding publication:

@inproceedings{di2026every,
title={Every Frame You Take: The Video-based Morphing Attack Detection Challenge},
author={Di Domenico, Nicol{`o} and Pellegrini, Lorenzo and La Cava, Simone Maurizio and Borghi, Guido and others},
booktitle={The IEEE International Joint Conference on Biometrics (IJCB)},
pages={1--6},
year={2026}
}