Official code repository for UPPET baseline training and evaluation.
Figure 1: Sample thermal scenes from the UPPET sub-datasets.
Figure 2: Keypoint annotations in diverse thermal scenes from the UPPET sub-datasets.
Human Pose Estimation (HPE) using visual images has achieved strong performance, but RGB-based models often degrade in low-light and adverse weather conditions. Thermal imaging offers a robust alternative and improved privacy.
UPPET unifies four thermal datasets under a common annotation protocol to support both specialization and cross-domain generalization studies in thermal HPE.
- Clone this repository.
- Create the conda environment.
- Prepare the official UPPET dataset:
https://huggingface.co/datasets/MCormier-KA/uppet_dataset
- Verify dataset availability:
python tools/check_dataset.py- Train a baseline:
python tools/train.py <config-file> --ampOne-shot setup:
bash tools/install_conda.shThe official UPPET dataset is hosted on Hugging Face:
https://huggingface.co/datasets/MCormier-KA/uppet_dataset
UPPET follows a split distribution model. The official dataset repository contains:
- redistributed TPE images
- UPPET annotations
- preparation scripts
- download utilities
- documentation
Prepare the dataset in the official repository before using this codebase.
Verify local structure in this repo:
python tools/check_dataset.pyExpected top-level paths in data:
- data/camel
- data/llvip
- data/otp
- data/tpe
- data/annotations
- configs: model and experiment configurations
- tools: training, testing, and dataset checks
- mmpose/datasets: dataset implementations used by this codebase
- runs: generated training logs and checkpoints
Train with a selected config:
python tools/train.py <config-file> --ampEvaluate a trained model:
python tools/test.py <config-file> <checkpoint-file>UPPET uses PoseTrack18 topology (15 keypoints) for all subsets. For MMPose compatibility, annotation and prediction files provide 17 keypoints, where keypoints 3 and 4 are ignored during metric computation.
- 2026-07-10 Official code release for UPPET baseline training and evaluation.
- 2026-07-09 Official UPPET dataset repository synchronized on Hugging Face: https://huggingface.co/datasets/MCormier-KA/uppet_dataset
- 2025-06-11 Paper accepted at the 21st Workshop on Perception Beyond the Visible Spectrum (PBVS 2025) at CVPR 2025.
Split-level overview for Specialization Case:
| 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 |
Dataset composition:
| Dataset | Images | Annotated Poses |
|---|---|---|
| TPE | 14,321 | 52,563 |
| CAMEL-P | 2,926 | 25,951 |
| LLVIP-Pose | 10,317 | 26,133 |
| OpenThermalPose | 6,090 | 14,277 |
| Split ID | LOO Training | LOO Evaluation | CV Training | CV Evaluation |
|---|---|---|---|---|
| 0 | OTP, CAMEL-P, TPE | LLVIP-Pose | LLVIP-Pose | OTP, CAMEL-P, TPE |
| 1 | LLVIP-Pose, CAMEL-P, TPE | OTP | OTP | LLVIP-Pose, CAMEL-P, TPE |
| 2 | LLVIP-Pose, OTP, TPE | CAMEL-P | CAMEL-P | LLVIP-Pose, OTP, TPE |
| 3 | LLVIP-Pose, CAMEL-P, OTP | TPE | TPE | LLVIP-Pose, CAMEL-P, OTP |
| Method | LLVIP-Pose | OTP | CAMEL-P | TPE | UPPET |
|---|---|---|---|---|---|
| HRNetw48-udp | 89.8 / 98.0 | 87.0 / 96.8 | 75.4 / 94.8 | 77.6 / 91.7 | 81.0 / 94.8 |
| ViTPose-s | 84.6 / 97.9 | 79.4 / 94.4 | 74.9 / 94.8 | 75.5 / 91.7 | 79.8 / 94.7 |
| ViTPose-h | 92.0 / 99.0 | 90.8 / 97.9 | 80.3 / 96.9 | 81.4 / 94.7 | 85.3 / 96.9 |
| SimCC | 85.6 / 98.0 | 81.8 / 95.7 | 72.0 / 93.6 | 73.6 / 90.5 | 78.7 / 93.7 |
| DeepPose-r50 | 80.7 / 96.9 | 76.9 / 93.8 | 66.6 / 92.7 | 69.9 / 87.4 | 76.2 / 92.6 |
| Method | Avg. | Split 0 | Split 1 | Split 2 | Split 3 |
|---|---|---|---|---|---|
| HRNetw48-udp | 54.9±12.8 / 78.0±11.6 | 65.2 / 88.5 | 51.8 / 78.3 | 35.3 / 59.0 | 67.3 / 86.3 |
| ViTPose-s | 49.3±14.1 / 75.9±13.1 | 62.7 / 88.1 | 40.0 / 75.8 | 31.2 / 54.5 | 63.3 / 85.1 |
| ViTPose-h | 68.3±13.4 / 85.4±12.1 | 81.4 / 96.7 | 73.0 / 91.3 | 45.9 / 65.0 | 72.9 / 88.6 |
| SimCC | 46.3±13.5 / 73.6±14.5 | 55.9 / 85.8 | 38.5 / 74.5 | 28.4 / 49.6 | 62.2 / 84.5 |
| DeepPose-r50 | 40.1±14.4 / 67.3±15.1 | 51.8 / 81.2 | 30.2 / 67.3 | 22.0 / 42.7 | 56.4 / 78.0 |
| Method | Avg. | Split 0 | Split 1 | Split 2 | Split 3 |
|---|---|---|---|---|---|
| HRNetw48-udp | 37.7±13.3 / 58.7±17.0 | 47.8 / 65.7 | 46.0 / 71.3 | 41.9 / 68.2 | 14.9 / 29.5 |
| ViTPose-s | 39.1±5.4 / 65.7±5.3 | 42.0 / 65.0 | 45.0 / 72.1 | 38.8 / 68.0 | 30.6 / 57.6 |
| ViTPose-h | 57.8±4.5 / 79.5±4.3 | 58.2 / 76.0 | 62.6 / 83.0 | 59.7 / 84.4 | 50.5 / 74.5 |
| SimCC | 32.0±12.0 / 54.7±17.0 | 39.7 / 59.4 | 37.8 / 63.4 | 39.4 / 70.0 | 11.2 / 26.1 |
| DeepPose-r50 | 27.6±10.6 / 50.5±14.3 | 38.1 / 58.9 | 34.4 / 58.7 | 27.2 / 58.6 | 10.5 / 25.7 |
If you use UPPET, please cite the UPPET paper and the original sub-dataset papers:
@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}
}This project is developed based on mmpose. We gratefully acknowledge the datasets and authors this work builds on: LLVIP-Pose, LLVIP, OpenThermalPose and CAMEL Dataset.