UDAPose

Official implementation of the paper

UDAPose: Unsupervised Domain Adaptation for Low-Light Human Pose Estimation, CVPR 2026

🛠️ Setup

Python environment

1. If you use uv, you can directly

   uv sync

   Or you can still use pip as following

   pip install -r requirements.txt

2. After that, you should compile and install the CUDA kernel for deformable attention.

   cd models/edpose/ops
   python setup.py build install

3. Test installation.

   python test.py

Prepare ExLPose dataset

1. Download ExLPose dataset and put it under data.

2. Organize the dataset as following:

   data
    |- ExLPose
        |- Annotations
        |   |- ExLPose_test_WL.json
        |   |- ExLPose-OC_test_A7M3.json
        |   |- ...
        |- bright
        |   |- (bright images)...
        |- dark
        |   |- (paired dark images)...
        |- ExLPose-OCN
        |   |- A7M3
        |   |   |- (A7M3 images)...
        |   |- RICOH3
        |   |   |- (RICOH3 images)...
   

🕵🏼 Inference

1. Download checkpoints from 🤗.

   hf download arsity/UDAPose-model-weights --local-dir ckpts

2. Run inference

   sh test.sh

You may also want to edit test.sh to evaluate on one subset.

🏋🏼 Train

Our full framework involves 3 steps in total. You can start from any step with our provided checkpoints or from the very beginning.

Train LCIM

1. Download SD 2.1 checkpoints from 🤗.

   hf download arsity/UDAPose-model-weights --local-dir ckpts

2. Start training LCIM

   cd data-pipeline
   accelerate launch train_vae.py

Results would be under ckpts/vae_train_outputs.

Generate training data

1. Download checkpoints from 🤗.

   hf download arsity/UDAPose-model-weights --local-dir ckpts

2. Start generating training data

   cd data-pipeline
   accelerate launch style_transfer.py

Synthetic training data would be under data/synthetic.

Train Pose Model

1. Download checkpoints from 🤗.

   hf download arsity/UDAPose-model-weights --local-dir ckpts

2. Download synthetic data from 🤗.

   hf download arsity/UDAPose-synthetic-data images.zip mapping_list.json --type dataset --local-dir data

3. Unzip and organize as following

   data
    |- mapping_list.json
    |- synthetic
        |- 0
        |- 1
        |- (image id directories)...
   

4. then

   sh train.sh

   to start training (for low-light). If you want to start from scratch (well-lit), you can edit train.sh.

©️ License

UDAPose is released under the Apache License 2.0 for our original contributions, unless otherwise noted.

This project builds upon several open-source projects. We preserve their original license notices, including:

• ED-Pose: Apache License 2.0, with the original ED-Pose license notices retained
• StyleID: MIT License

Some optional components, pretrained models, or external checkpoints may be subject to their own licenses, such as Stable Diffusion / Swin-Transformer. These are not covered by the Apache-2.0 license of our original code.

Please see LICENSES/ for details.

📝 Bibtex

If you find this work useful, please consider cite our paper

@InProceedings{chen2026udapose,
    author    = {Chen, Haopeng and Ai, Yihao and Kim, Kabeen and Tan, Robby T. and Chen, Yixin and Wang, Bo},
    title     = {UDAPose: Unsupervised Domain Adaptation for Low-Light Human Pose Estimation},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2026},
    pages     = {13781-13792}
}