MCormier-KA/uppet_dataset

UPPET Dataset

UPPET (Unified Pedestrian Pose Estimation in Thermal Imaging) is a benchmark for pedestrian pose estimation in thermal imagery. It unifies four thermal datasets under a common annotation format and evaluation protocol, enabling both within-dataset and cross-dataset evaluation of human pose estimation methods.

The benchmark combines the newly introduced CAMEL-P and TPE datasets with LLVIP-Pose and an extended version of OpenThermalPose, providing 33,654 thermal images with 118,924 annotated pedestrian poses across diverse thermal sensors, image resolutions, and real-world scenarios. UPPET is the first benchmark specifically designed to evaluate the generalization ability of human pose estimation methods across multiple thermal datasets.

This repository serves as the official distribution of the UPPET benchmark. It contains all UPPET annotations, the redistributable TPE images, and standalone preparation scripts that automatically assemble the complete benchmark while respecting the licenses of the original datasets.

Dataset Overview

Split	Images	Annotated Poses	Avg. Keypoints / Pose	Avg. Poses / Image
Train	23,291	86,316	12.29	3.91
Test	10,363	32,608	12.08	3.48
Total	33,654	118,924	12.23	3.53

The UPPET benchmark is composed of four thermal datasets:

Dataset	Images	Annotated Poses	Included in this repository
TPE	14,321	52,563	✅
CAMEL-P	2,926	25,951	❌
LLVIP-Pose	10,317	26,133	❌
OpenThermalPose	6,090	14,277	❌

Only the TPE images are redistributed in this repository. Images from CAMEL-P, LLVIP-Pose, and OpenThermalPose remain under their original licenses and are automatically downloaded or prepared by the included preparation pipeline.

📄 Paper: https://openaccess.thecvf.com/content/CVPR2025W/PBVS/html/Cormier_UPPET_Unified_Pedestrian_Pose_Estimation_in_Thermal_Imaging_CVPRW_2025_paper.html

💻 Code: https://github.com/MickaelCormier/uppet

---

News

• 2026-07-09 — UPPET dataset re-uploaded and synchronized on Hugging Face.
• 2025-06-11 — Our paper UPPET: Unified Pedestrian Pose Estimation in Thermal Imaging was accepted at the 21st Workshop on Perception Beyond the Visible Spectrum (PBVS 2025) at CVPR 2025.

---

Licensing Model and Data Distribution

UPPET follows a split distribution model to respect the licenses of the original datasets.

Component	Included	Description
TPE images	✅	Redistributed in this repository
UPPET annotations	✅	COCO-style annotations for the complete benchmark
CAMEL-P images	❌	Downloaded from the original source
LLVIP images	❌	Downloaded from the original source
OpenThermalPose images	❌	Downloaded from the original source

Redistributed assets:

• TPE images are distributed in data/tpe/images.
• UPPET annotations are distributed in data/annotations.

Images from CAMEL-P, LLVIP, and OpenThermalPose are not redistributed in this repository. They remain under their original licenses and must be obtained from their respective sources.

Redistributed assets are released under the CC BY-NC-SA 3.0 license.

The preparation pipeline reflects this distribution model:

• validates the locally distributed TPE images,
• validates the distributed UPPET annotations,
• downloads or prepares CAMEL-P, LLVIP, and OpenThermalPose from their official sources.

---

Repository Structure

Included in this repository:

• data/annotations: UPPET COCO-style annotations (CV, LOO, specialization)
• data/tpe/images: TPE images redistributed with UPPET
• tools/prepare_data: standalone preparation scripts

After running the preparation pipeline, the following directories will additionally be created:

• data/camel/images
• data/llvip/infrared
• data/otp/images

---

Setup

Install the preparation dependencies:

pip install -r tools/prepare_data/requirements.txt

If the repository was cloned with Git LFS pointers only, download the tracked files:

git lfs install
git lfs pull

---

Prepare the Full UPPET Dataset

Run from the repository root:

python tools/prepare_data/prepare_data.py

The preparation script automatically:

• prepares CAMEL-P images in data/camel/images,
• prepares LLVIP images in data/llvip/infrared,
• prepares OpenThermalPose images in data/otp/images,
• validates local TPE images in data/tpe/images,
• validates local UPPET annotations in data/annotations.

The preparation pipeline is fully resumable:

• already prepared datasets are skipped,
• existing local archives (CAMEL-P, LLVIP, OpenThermalPose) are reused instead of being downloaded again.

---

Individual Scripts

• tools/prepare_data/download_camel.py
• tools/prepare_data/download_llvip.py
• tools/prepare_data/download_otp.py
• tools/prepare_data/download_tpe.py
• tools/prepare_data/download_annotations.py

---

Human Pose Topology

UPPET uses the PoseTrack18 topology with 15 annotated keypoints for all subsets.

For compatibility with MMPose, annotation and prediction files provide 17 keypoints, where keypoints 3 and 4 are ignored during metric computation.

---

Citation

If you use the UPPET dataset in your research, please cite the UPPET paper together with the original datasets included in the benchmark.

@InProceedings{Cormier_2025_CVPR,
    author    = {Cormier, Mickael and Specker, Andreas and Beyerer, J\"urgen},
    title     = {UPPET: Unified Pedestrian Pose Estimation in Thermal Imaging},
    booktitle = {Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR) Workshops},
    month     = {June},
    year      = {2025},
    pages     = {4551-4560}
}

@InProceedings{Cormier_2024_ACCV,
    author    = {Cormier, Mickael and Ng Zhi Yi, Caleb and Specker, Andreas and Bla{\ss}, Benjamin and Heizmann, Michael and Beyerer, J{\"u}rgen},
    title     = {Leveraging Thermal Imaging for Robust Human Pose Estimation in Low-Light Vision},
    booktitle = {Proceedings of the Asian Conference on Computer Vision (ACCV) Workshops},
    month     = {December},
    year      = {2024}
}

@InProceedings{jia2021llvip,
    author    = {Jia, Xinyu and Zhu, Chuang and Li, Minzhen and Tang, Wenqi and Zhou, Wenli},
    title     = {LLVIP: A visible-infrared paired dataset for low-light vision},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision},
    pages     = {3496--3504},
    year      = {2021}
}

@InProceedings{kuzdeuov2014,
    author    = {Kuzdeuov, Askat and Taratynova, Darya and Tleuliyev, Alim and Varol, Huseyin Atakan},
    title     = {OpenThermalPose: An Open-Source Annotated Thermal Human Pose Dataset and Initial YOLOv8-Pose Baselines},
    booktitle = {2024 18th IEEE International Conference on Automatic Face and Gesture Recognition (FG)},
    year      = {2024},
    doi       = {10.1109/FG59268.2024.10581992}
}

@InProceedings{gebhardt2018camel,
    author    = {Gebhardt, Evan and Wolf, Marilyn},
    title     = {Camel dataset for visual and thermal infrared multiple object detection and tracking},
    booktitle = {2018 15th IEEE International Conference on Advanced Video and Signal Based Surveillance (AVSS)},
    year      = {2018}
}

---

Acknowledgment

This preparation pipeline and dataset harmonization build upon the following publicly available datasets and toolchains:

• LLVIP-Pose: https://github.com/MickaelCormier/llvip-pose/
• LLVIP: https://github.com/bupt-ai-cz/LLVIP
• OpenThermalPose: https://github.com/IS2AI/OpenThermalPose/
• CAMEL Dataset: https://camel.ece.gatech.edu/

We gratefully acknowledge the authors of these datasets for making their work publicly available and enabling research on thermal human pose estimation.