Retrained ViTPose Models

This dataset contains five retrained ViTPose-large models.
They were generated for
Jahn, L., Flügge, S., Zhang, D. et al. Comparison of marker-less 2D image-based methods for infant pose estimation. Sci Rep 15, 12148 (2025). https://doi.org/10.1038/s41598-025-96206-0
Please note that the performance of these models will depend on the specificity
of your dataset. Please see the publication for more details on this matter.

To run a model, install mmpose (https://github.com/open-mmlab/mmpose/tree/main)
and configure everything as if you wanted to use ViTPose-large
(https://mmpose.readthedocs.io/en/latest/model_zoo/body_2d_keypoint.html#
topdown-heatmap-vitpose-on-coco), then swap out the .pth file for one provided
here and change the filename accordingly.

Dataset Card for MPII Human Pose Descriptions

  Dataset Summary

The MPII Human Pose Descriptions dataset extends the widely-used MPII Human Pose Dataset with rich textual annotations. These annotations are generated by various state-of-the-art language models (LLMs) and include detailed descriptions of the activities being performed, the count of people present, and their specific poses. The dataset consists of the same image splits as provided in MMPose, with 14644… See the full description on the dataset page: https://huggingface.co/datasets/saifkhichi96/mpii-human-pose-captions.

COCO-WholeBody is an extension of COCO dataset with whole-body annotations. There are 4 types of bounding boxes (person box, face box, left-hand box, and right-hand box) and 133 keypoints (17 for body, 6 for feet, 68 for face and 42 for hands) annotations for each person in the image.

Advances in machine learning and wearable sensors offer new opportunities for capturing and analyzing human movement outside specialized laboratories. Accurately tracking and evaluating human movement under real-world conditions is essential for telemedicine, sports science, and rehabilitation. This work introduces a comprehensive benchmark comparing deep learning monocular video-based human pose estimation models with inertial measurement unit (IMU)-driven methods, leveraging VIDIMU dataset containing a total of 13 clinically relevant activities which were captured using both commodity video cameras and 5 IMUs. Joint angles derived from state-of-the-art deep learning frameworks (MotionAGFormer, MotionBERT, MMPose 2D-to-3D pose lifting, and NVIDIA BodyTrack included in Maxine-AR-SDK) were evaluated against joint angles computed from IMU data using OpenSim inverse kinematic methods. A graphical comparison of the angles estimated by each model shows the overall performance for each activity.

The results, which also contains the evaluation of multiple metrics (RMSE, NMRSE, MAE, correlation and coefficient of determination) in table and plot format, highlight key trade-offs between video- and sensor-based approaches including costs, accessibility and precision across different daily life activities. This work establishes valuable guidelines for researchers and clinicians seeking to develop robust, cost-effective, and user-friendly solutions for telehealth and remote patient monitoring solutions, ultimately bridging the gap between AI-driven motion capture and accessible healthcare applications.