Human Pose Estimation

Human motion capture is defined as the process of recording the movement of a person from sensor measurements. For decades video-based MoCap has been realized by placing optical markers on the human body and capture them with several synchronized cameras. While such systems achieve high accuracies they are impractical to apply on a daily basis, for example to improve athletes’ performances, assist with rehabilitation, or enabling human machine interaction.

We simplify motion capture for everybody, mostly requiring only a single camera, by developing machine learning based techniques to accurately capture 3-dimensional human motion. Since this process strongly depends on high quality training data, and such data is at best sparsely available, we have a special focus on weakly supervised learning techniques. 

The publications below give an overview of our latest advances.


Mirror-aware Neural Humans
Daniel Ajisafe, James Tang, Shih-Yang Su, Bastian Wandt, Helge Rhodin
3DV 2023

DiffPose: Multi-hypothesis Human Pose Estimation using Diffusion models,
Karl Holmquist, Bastian Wandt
ICCV 2023

ElePose: Unsupervised 3D Human Pose Estimation by Predicting Camera Elevation and Learning Normalizing Flows on 2D Poses,
Bastian Wandt, Jim Little, Helge Rhodin
CVPR 2022

AdaptPose: Cross-Dataset Adaptation for 3D Human Pose Estimation by Learnable Motion Generation,
Mohsen Gholami, Bastian Wandt, Helge Rhodin, Rabab Ward, Z. Jane Wang
CVPR 2022

Probabilistic Monocular 3D Human Pose Estimation with Normalizing Flows,
Tom Wehrbein, Marco Rudolph, Bodo Rosenhahn, Bastian Wandt
ICCV 2021
[pdf] [code]

CanonPose: Self-supervised Monocular 3D Human Pose Estimation in the Wild
Bastian Wandt, Marco Rudolph, Helge Rhodin, Bodo Rosenhahn
CVPR 2021
[pdf] [code]