SynThermFace

Description

SynThermFace-Digi2Real is a synthetic heterogeneous face recognition dataset consisting of paired visible (VIS) and synthetic thermal face images for 1000 synthetic identities. For each selected visible image from the Digi2Real dataset, a corresponding synthetic thermal image is generated using the SynThermFace VIS-to-thermal synthesis pipeline. The thermal generation model is based on Qwen-Image-Edit and is adapted using paired real VIS–thermal face images from the MCXFace dataset. The dataset is intended primarily for training and adapting heterogeneous face recognition models across the visible and thermal domains.

The SynThermFace-Digi2Real dataset is derived from the Digi2Real synthetic face dataset, which provides photorealistic synthetic visible-spectrum face images with identity labels. A subset of 1000 Digi2Real identities is selected, and the corresponding visible images are used as inputs to the thermal generation pipeline introduced in SynThermFace. The generation model is based on Qwen-Image-Edit and is adapted for VIS-to-thermal face translation using paired visible and thermal images from MCXFace. MCXFace provides spatially and temporally registered real RGB and thermal face images and is introduced in the Prepended Domain Transformer work for heterogeneous face recognition.

For every generated thermal image, the corresponding Digi2Real visible image is also included in the dataset, resulting in explicitly paired VIS–thermal samples. The two modalities are stored in separate `VIS` and `THERMAL` directories, with one subdirectory per synthetic identity. The filenames preserve the correspondence between modalities. For example, the visible image `1_image_1.jpg` corresponds to the synthetic thermal image `1_image_1_nir.jpg`. Thus, the corresponding visible image can be obtained by removing the `_nir` suffix from the thermal filename.

All samples are stored as `.jpg` images. The identities and visible images are synthetic, and the thermal images are also synthetically generated rather than acquired with a physical thermal sensor. SynThermFace-Digi2Real is therefore intended mainly as a training and adaptation resource for VIS–thermal heterogeneous face recognition, cross-spectral representation learning, domain adaptation, and related synthetic-data-based biometric research. Evaluation of models trained using this dataset should preferably be performed on independent datasets containing real thermal imagery.

Reference

The SynThermFace dataset was originally introduced in the following paper:

@inproceedings{george2026synthermface,
  title={SynThermFace: Amplifying Limited Paired Data for Visible--Thermal Face Recognition via Synthetic Data Generation},
  author={George, Anjith and Unal, Adam and Marcel, S\'ebastien},
  booktitle={British Machine Vision Conference (BMVC)},
  year={2026},
}

If you use the dataset, please consider citing the following references for Digi2Real and MCXFace as well.

@inproceedings{george2025digi2real,
  title={Digi2real: Bridging the realism gap in synthetic data face recognition via foundation models},
  author={George, Anjith and Marcel, S{\'e}bastien},
  booktitle={2025 IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW)},
  pages={1379--1388},
  year={2025},
  organization={IEEE}
}

@article{george2022prepended,
  title={Prepended domain transformer: Heterogeneous face recognition without bells and whistles},
  author={George, Anjith and Mohammadi, Amir and Marcel, Sebastien},
  journal={IEEE transactions on information forensics and security},
  volume={18},
  pages={133--146},
  year={2022},
  publisher={IEEE}
}