RGBT

Introduction

The code implementation for the paper:Towards Privacy-Preserving Thermal Human Perception from Dataset to Deployment.

Installation

Download or clone the repository.

git clone https://github.com/kailaisun/RGBT.git
cd RGBT

Environment Installation

We recommend using Conda (Miniconda) for installation.

Dataset summary

Some examples:

We publish part of the thermal images and annotations on Hugging Face.

Deployment

Released models

We release one model per task, together with minimal inference code. Weights live on Hugging Face; each folder below contains a self-contained infer.py.

Task	Method	Folder	Weights (Hugging Face)	Headline metric
Infrared person counting	ResNet18	resnet18/	resnet18/weights/best.pt	accuracy / macro-F1
RGB to thermal field	U-Net	unet_rgb2t/	unet_rgb2t/weights/checkpoint.pt	MAE / R2
Infrared to RGB	BBDM	bbdm_ir2rgb/	bbdm_ir2rgb/weights/last_model.pth	PSNR / SSIM / FID

Only the thermal-infrared image is used as input for the counting model; RGB is never fed to it.

The infrared person-state detection model (4 states: lie / sit / other / off_bed) is coming soon.

Download the weights

pip install "huggingface_hub[cli]"
hf download skl24/RGBT --local-dir checkpoints

Run inference

# person counting
python resnet18/infer.py  --weights checkpoints/resnet18/weights/best.pt           --source ir.png
# RGB -> thermal field
python unet_rgb2t/infer.py --weights checkpoints/unet_rgb2t/weights/checkpoint.pt  --source rgb.jpg
# IR -> RGB
python bbdm_ir2rgb/infer.py --weights checkpoints/bbdm_ir2rgb/weights/last_model.pth \
    --source ir.png --output rgb.png

See each folder's README.md for the model details, training setup and held-out test numbers.

Citation

License

The repository is licensed under the MIT license.

Contact Us

If you have other questions❓, please contact us in time 👬