New research from The Podium Institute uses generative AI to create realistic movement data that could support future advances in sports science, rehabilitation and human health.
Understanding how people move is important for everything from improving sporting performance to supporting recovery from injury. Analysing movement accurately requires a huge amount of high-quality data and, until now, this has been one of the biggest barriers to progress in this area.
Researchers from The Podium Institute for Sports Medicine and Technology at the University of Oxford and the Visual Geometry Group are directly tackling this challenge with PoseDreamer.
By generating a large dataset of 500,000 synthetic images, PoseDreamer could help researchers train movement-analysis technologies at scale, reducing reliance on collecting large volumes of real-world participant data.
Overcoming one of the biggest challenges in movement research
Collecting movement data is often expensive and time-consuming, relying on specialist motion-capture laboratories and equipment. PoseDreamer offers a new approach, generating realistic training data at scale while maintaining detailed information about body position and movement – as presented in the paper PoseDreamer: Scalable and Photorealistic Human Data Generation Pipeline with Diffusion Models.
Commenting on the findings, Lorenza Prospero, first author of the paper, said:
“PoseDreamer is a first step towards bridging the gap between the realism of real-world photographs and the precise 3D labelling of computer-rendered images. Across diverse models and domains, the two approaches have strengths and weaknesses which can complement and enhance one another."
Creating tailored data for sport and beyond
The researchers also showed that the system can be adapted for specific activities. For example, generating more yoga-focused examples improved performance on yoga movements, highlighting the potential to create tailored datasets for different sports and applications.
By creating realistic synthetic movement data, PoseDreamer could help researchers overcome two longstanding challenges: obtaining enough high-quality data and reducing reliance on large collections of real-world participant images. This opens new opportunities to develop movement-analysis technologies at a scale that would previously have been difficult or costly to achieve.