Robust Long-Term Test-Time Adaptation for 3D Human Pose Estimation Through Motion Discretization

Official implementation of the paper by Yilin Wen, Kechuan Dong, and Yusuke Sugano: "Robust Long-Term Test-Time Adaptation for 3D Human Pose Estimation Through Motion Discretization", AAAI 2026.

[Paper and Appendix] | [Supplementary Video] | [Project Page]

If you find this work helpful, please consider citing:

@inproceedings{wen2026robust,
  title={Robust Long-Term Test-Time Adaptation for 3D Human Pose Estimation Through Motion Discretization},
  author={Wen, Yilin and Dong, Kechuan and Sugano, Yusuke},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  year={2026}
}

Requirements and Preprocessing

Environment

The code is tested with the following environment:

Ubuntu 22.04
Python 3.10
PyTorch 2.4.0

You can create a conda environment for this project by running:

conda env create -f environment.yml
conda activate OTTPA

If you encounter issues with torchgeometry, you may modify the kernel code by referring to I2L-Mesh or CycleAdapt.

Pretrained Models and SMPL Related Files

Please prepare the pretrained models and other SMPL-related files in the ./assets/ directory following this structure:

${root_dir_of_this_repo}/
├── assets/
│   ├── human_models/
│   │   ├── J_regressor_h36m_smpl.npy
│   │   ├── smpl_mean_params.npz
│   │   └── smpl/
│   │       ├── SMPL_FEMALE.pkl
│   │       ├── SMPL_MALE.pkl
│   │       └── SMPL_NEUTRAL.pkl
│   ├── pose_prior/
│   │   └── gmm_08.pkl
│   └── pretrained_models/
│       ├── hmr_basemodel.pt
│       ├── motionnet.pth
│       └── motionnet_stats/
│           ├── mean.pt
│           └── std.pt
├── configs/
├── core/
├── dataset/
├── exp/
├── models/
├── utils/
└── ...

All files except those under smpl/ can be downloaded from: Link.

For SMPL models, please download them from the official SMPL website.

Data Preprocessing

Download Datasets

Please download 3DPW and Ego-Exo4D from their official websites.

Download 2D Detection and Miscellaneous Files

We provide precomputed 2D detections (OpenPose and ViTPose), as well as other miscellaneous files at: Link.

Save Images to LMDB

We save images to LMDB for adaptation using convert_to_lmdb.py:

python convert_to_lmdb.py --cfg <path_to_cfg> --participant_uid <uid>

Parameters:

• --cfg: We provide two example configuration files under the configs directory for running on the 3DPW and Ego-Exo4D datasets. While you can ignore most parameters at this step, you need to properly set:

  • dir_root: root directory of the downloaded dataset
  • dir_lmdb: root directory of the output image LMDB
  • dir_misc: root directory of 2D detection and miscellaneous files obtained

• --participant_uid:

  • For Ego-Exo4D dataset, it indicates the participant UID to be processed
  • For 3DPW dataset, this parameter will be ignored at this stage, as images of all participants will be processed in a single run

Examples:

For 3DPW subjects:

python convert_to_lmdb.py --cfg ./configs/3dpw.yml

For an Ego-Exo4D subject:

python convert_to_lmdb.py --cfg ./configs/egoexo.yml --participant_uid 161

Running Experiments

Adaptation and Evaluation

Use adapt.py to run adaptation on input videos:

python adapt.py --gpu 0 --cfg <path_to_config> --participant_uid <uid>

Parameters:

• --cfg: In the configuration file, you also need to properly indicate the 2D joint detections (vitpose or openpose) used for adaptation.

• --participant_uid:

  • On Ego-Exo4D, we test with participants with UIDs: 388, 386, 387, 384, 383, 323, 318, 376, 320, 316, 382, 558, 317, 161, 549, 525, 544, 34, 40, 244, 243, 245, 239, 240, 46, 226, 24, 254, 227, 256 (as reported in Table 5 of our appendix).
  • For 3DPW, use values 0-4, indicating the five participants involved in the test set.

Examples:

For a 3DPW subject:

python adapt.py --gpu 0 --cfg ./configs/3dpw.yml --participant_uid 1

For an Ego-Exo4D subject:

python adapt.py --gpu 0 --cfg ./configs/egoexo.yml --participant_uid 161

Result Visualization

Use visualize_result.py to visualize estimation results after adaptation. This will produce figures similar to Figure 3 in our main text.

python visualize_result.py --dir_exp <dir_to_exp>

Parameters:

• --dir_exp: Root directory of experiment results, which is the cfg.output_dir used in adaptation. The script will load the log file and estimation results saved during adaptation, and save visualization results under <dir_to_exp>/results.

License and Acknowledgement

Part of the codebase relies on the following repositories:

• CycleAdapt, ICCV 2023
• STMC, CVPRW 2024
• T2M-GPT, CVPR 2023
• MoMask, CVPR 2024

The pretrained MotionNet (motionnet.pth) is under the ModelGo Attribution-NonCommercial-ResponsibleAI License, Version 2.0 (MG-NC-RAI-2.0).