MAli-Farooq/ThermVision-DB

🥽 ThermVision-DB: Synthetic Thermal Face Dataset

Author: Muhammad Ali Farooq Affiliation: C3I Group, School of Engineering, University of Galway License: Research-only, non-commercial Modality: Thermal (LWIR) Face Images + Videos Resolution: 512 × 512 Version: 1.0

🧩 Dataset Summary

ThermVision-DB is a synthetic thermal facial dataset generated using thermally tuned FLUX.Dev diffusion-based models and Image-to-Video Translation methods specifically for long-wave IR imaging. The dataset supports research in thermal biometrics, privacy-preserving identity modelling, thermal domain adaptation, and multimodal fusion.

The dataset includes high-quality thermal facial images and controlled thermal video sequences with variations in pose, expression, accessories, and gender representation. All samples are synthetic and contain no personally identifiable information.

🔥 Key Features

Fully synthetic & privacy-preserving

High-resolution thermal faces (512×512)

Balanced gender representation (50 subjects: 25 male, 25 female)

Attribute-rich content: head pose, expressions, accessories, hairstyles

Paired video sequences for each subject

Annotations for gender, identity, and face bounding boxes

Supports training of thermal recognition, domain adaptation, and generative models

📊 Dataset Structure

ThermVision-DB Video Data

Field	Details
Total Subjects / Gender	50 (25 male, 25 female) — extendable to unlimited synthetic subjects
Format & Resolution	MP4, 512×512
Per-Subject Data	2 video sequences, 604 frames total
Total Dataset Size	100 videos, ~30.2K frames
Annotations	Gender, identity labels, face bounding boxes
Attributes	Head poses, expressions, hairstyles, accessories, gender parity

ThermVision-DB 2D Facial Image Data

Category	Description / Content Type	Resolution	Gender
Beard	Thermal facial images featuring subjects with beards.	512×512	Male
Facial Accessories	Includes eyewear, masks, hats, or other accessories.	512×512	Male / Female
Frontal Pose	Frontal-facing thermal face images of various subjects.	512×512	Male / Female
Mix	Mixed data combining poses, expressions, and variations.	512×512	Male / Female
Poses	Thermal face images with varying head poses and orientations.	512×512	Male / Female

🧠 Generation Pipeline

Thermal Diffusion Model: Fine-tuned using multi-seed thermal datasets and LoRA modules for identity-preserving synthesis.

Attribute-controlled prompts: Used for generating poses, expressions, accessories, and gender-balanced samples.

Video Generation: Thermal frame interpolation + controlled identity-consistent animation pipeline.

Annotation Pipeline: Automatically extracted face detector bounding boxes + metadata tagging (gender, identity, attributes).

🎯 Intended Use

This dataset is designed for:

1. Thermal face recognition

2. Cross-spectral domain adaptation (RGB ↔ thermal)

3. Privacy-preserving biometrics

4. Expression & pose analysis in thermal domain

5. Robust perception for surveillance & low-light environments

6. Generative modelling & diffusion research

🔐 Ethics & License

All samples are synthetic, contain no real identities, and are generated under a strictly controlled pipeline following FAIR, GDPR, and privacy-preserving design principles.

---

License: cc-by-nc-4.0

This dataset is licensed under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license. This license allows for non-commercial use of the dataset, provided proper attribution is given to the authors. Adaptations and modifications are permitted for non-commercial purposes. For full details, please review the CC BY-NC 4.0 license and the license file included in the dataset. As the dataset is gated, you must accept the access conditions on Hugging Face to use it.

✔️ Free for academic and research use ❌ Commercial use prohibited without written permission ❌ Redistribution not allowed

📜 Citation

If you use ThermVision-DB in your research, please cite our dataset and paper

@inproceedings{farooq2025thermvision, title={ThermVision: Exploring FLUX for Synthesizing Hyper-Realistic Thermal Face Data and Animations via Image to Video Translation}, author={Farooq, Muhammad Ali and Shariff, Waseem and Corcoran, Peter}, booktitle={Proceedings of the 33rd ACM International Conference on Multimedia}, pages={10161--10170}, year={2025}