MLAdaptiveIntelligence/PRIMA

MODEL CARD:

Model Details

• PRIMA model(s) developed by the M.W.Mathis Lab in 2026, trained to predict quadruped shape and pose from images. Please see paper link for details.

• There are two main models:

• s1ckpt.ckpt is the stage-1 model trained with Animal3D, CtrlAni3D, and Quadruped2D datasets.
• s3ckpt.ckpt is the stage-3 model trained with Animal3D, CtrlAni3D, and Quadruped3D datasets.

from pathlib import Path
from huggingface_hub import hf_hub_download

repo_id = "MLAdaptiveIntelligence/PRIMA"

model_dir = Path("./prima_model")
model_dir.mkdir(parents=True, exist_ok=True)

# download stage-1 checkpoint
s1_path = hf_hub_download(
    repo_id=repo_id,
    filename="s1ckpt.ckpt",
    local_dir=model_dir
)

# donwload stage-3 checkpoint
s3_path = hf_hub_download(
    repo_id=repo_id,
    filename="s3ckpt.ckpt",
    local_dir=model_dir
)

Intended Use

• Intended to be used for shape and pose estimation of quadruped images taken from a single view.

• Intended for academic and research professionals working in fields related to animal behavior, such as neuroscience and ecology.

• Not suitable as a zero-shot model for applications that require high shape and pose precision, but can be further optimized with 2D keypoint annotations or from SuperAnimal to improve accuracy. Also, it is not suitable for videos that look dramatically different from those we show in the paper.

Metrics

• PA-MPJPE (Procrustes-aligned mean per-joint position error), computed over 3D joints.

• PA-MPVPE (Procrustes-aligned mean per-vertex position error), computed over the SMAL mesh vertices.

• PCK (Percentage of Correct Keypoints) measures the proportion of predicted keypoints within a specified threshold of the ground-truth keypoints.

• AUC (Area Under the Curve), computed by integrating the PCK values as the threshold varies from 0 to 1.

Evaluation Data

• In the paper we benchmark on Animal3d, CtrlAni3D, Quadruped2D, and AnimalKingdom.

Training Data:

It consists of being trained together on the following datasets:

• Animal3D see full details at (1).
• CtrlAni3D See full details at (2).
• Quadruped2D See full details at (3).
• Quadruped3D See full details at paper link.

Ethical Considerations

• No experimental data were collected for this model; all datasets used are cited.

License

Modified MIT.

Copyright 2026 by Mackenzie Mathis, Xiaohang Yu, and contributors.

Permission is hereby granted to you (hereafter "LICENSEE") a fully-paid, non-exclusive, and non-transferable license for academic, non-commercial purposes only (hereafter “LICENSE”) to use the "MODEL" weights (hereafter "MODEL"), subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software:

This software may not be used to harm any animal deliberately.

LICENSEE acknowledges that the MODEL is a research tool. THE MODEL IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE MODEL OR THE USE OR OTHER DEALINGS IN THE MODEL.

If this license is not appropriate for your application, please contact Prof. Mackenzie W. Mathis (mackenzie@post.harvard.edu) and/or the TTO office at EPFL (tto@epfl.ch) for a commercial use license.

Please cite paper link if you use this model in your work.

References

1. Xu, J., Zhang, Y., Peng, J., Ma, W., Jesslen, A., Ji, P., Hu, Q., Zhang, J., Liu, Q., Wang, J., et al.: Animal3d: A comprehensive dataset of 3d animal pose and shape. In: ICCV. pp. 9099–9109 (2023)
2. Lyu, J., Zhu, T., Gu, Y., Lin, L., Cheng, P., Liu, Y., Tang, X., An, L.: Animer: Animal pose and shape estimation using a family-aware transformer. In: CVPR. pp. 17486–17496 (2025)
3. Ye, S., Filippova, A., Lauer, J., Schneider, S., Vidal, M., Qiu, T., Mathis, A., Mathis, M.W.: Superanimal pretrained pose estimation models for behavioral analysis. Nature communications 15(1), 5165 (2024)