On March 19th, 2019, CloudWalk (云从科技) claimed that it made breakthroughs in 3D human body reconstruction and created new recordS in three largest datasets in 3D human post estimation Human 3.6M, SURREAL, and UP-3D.
CloudWalk, an AI startup hatched by Chinese Academy of Sciences (CAS), is a unicorn valued at USD 3.54 billion according to itjuzi from its latest financing series in October 2018. As one of the CV (computer vision) unicorns, CloudWalk’s team is backed up by outstanding CV scientists and engineers with the CAS heritage.
In 3D human pose estimation, the error is the critical index to justify whether the 3D human pose estimation algorithm’s performance. The lower the error, the finer in accuracy and better in performance. CloudWalk’s 3D human pose estimation’s minimum error is decreased by approximately 30%.
From the above comparisons within different datasets, CloudWalk’s 3D human pose estimation algorithm presents better performance than previously-known algorithms. CloudWalk is reputable in the facial recognition area and it provided the first prototype for instant payment based on facial recognition. The application of 3D human pose estimation can be used in cloth parsing, action recognition, human-object interaction, and pose search in database as introduced by Digvijay Singh.
With CAS’ support, CloudWalk’s business deeply roots in the finance area, and with great potential in security as well. Action recognition in the security area is less popular than facial recognition because it is not widely used for identification purpose. Comparing facial recognition’s powerful function in the identification, action recognition serves more to predict and identify possible dangerous and even terrorist behaviors based on machine learning. Besides security purpose, action recognition can be used for the physical-related purpose such as physical training and analysis.
CloudWalk’s improvement in accuracy can provide better 3D human body reconstruction. The algorithm that CloudWalk invented has lower requirements for input image while maintaining accuracy. The lowered requirement marks that general optical devices can become the image input resource and make 3D pose estimation available to more individuals.
Not only the loosen requirements on images, the quantity of images required to perform 3D pose estimation is also reduced. CloudWalk can employ only one frame of the image to conduct 3D human body reconstruction. Reduced quantity and lowered quality of images can introduce 3D pose estimation into more applications where these applications might have restrictions in either quality or quantity of images.
To conduct 3D human body reconstrction with less-preferred quality and quantity in the image, it requires the algorithm to be much more powerful. To build a 3D image based on 2D image needs the algorithm to have accurate estimates based on perspectives, shade overlay, and other optical principles to analyze the 3D feature point’s position and direction. The complexity embedded in algorithms enables more images can be used for 3D human pose estimation from a broader range.