By Peng Li, EDITOR
On March 19, 2019, CloudWalk claimed breakthroughs in 3D human body reconstruction technology based on one frame images and new world records in the three largest datasets in Human 3.6M, SURREAL, and UP-3D. This greatly reduces the original minimum error by 30%. It is also another important achievement for CloudWalk’s in 3D human body reconstruction technology, following its record-breaking 3D facial datasets last year.
Performance Comparison on Human3.6M
Performance Comparison on Surreal
Performance Comparison on UP-3D
As for 3D human body reconstruction technology, the error is the critical index to justify whether the 3D human pose estimates the algorithm’s performance. The lower the error, the finer in accuracy and better in performance.
Above the tables, the surface error is reduced to 52.7mm from 75.4mm, followed by the 3D joint error that is reduced from 55.8mm to 40.1mm on Surreal. As for Human3.6M, the 3D joint error is decreased to 46.7mm from 59.9mm. The execution speed descended from hundreds of milliseconds to only 5 milliseconds.
According to Tian Guodong, a senior algorithm researcher at CloudWalk, the technology can be applied in such fields as medical simulation limb printing, virtual fitting, facial makeup and expressive facial animation synthesis.
It is worth mentioning that Human3.6M, Surreal and UP-3D are the global authoritative datasets on 3D human body reconstruction technology. Some renowned enterprises, institutes and universities including University of California, Berkeley, Max Planck Institute, Amazon, University of Pennsylvania, Peking University, Zhejiang University, Zhejiang University, French Institute for Research in Computer Science and Automation and Adobe Research, are all engaged in fierce competition for algorithm capability in this regard. Compared with the past, Chinese enterprises and universities are beginning to show their advantages in foreign established fields.