UDAPose

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

Roadmap

• Inference code and model weights
• Pose model training and synthetic data
• Data synthesis pipeline and training

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

Inference

1. Download checkpoints from 🤗.

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

2. Download ExLPose dataset and put into data.

3. 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)...
   

4. Run inference

   sh test.sh

Train

Train Pose Model

Download synthetic data from 🤗.

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

Unzip and organize as following

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

then

sh train.sh

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

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}
}