Humans construct AI techniques to take care of environments largely populated by our personal variety, and so it’s not shocking that one of many extra common pc imaginative and prescient analysis areas is in human pose estimation. In the latest paper Contact and Human Dynamics from Monocular Video, a analysis workforce from Stanford University and Adobe Research proposes a brand new strategy that mixes discovered pose estimation with bodily reasoning by way of trajectory optimization to extract dynamically legitimate full-body motions from monocular video. The researchers say the strategy produces motions which are visually and bodily way more believable than state-of-the-art strategies.
Existing strategies for human pose estimation from monocular video can estimate 2D and 3D kinematic poses. These strategies nevertheless typically nonetheless comprise seen errors that defy bodily constraints, such because the toes from recovered motions for instance floating barely above or penetrating into the bottom. These errors can then distort or stop subsequent makes use of of the movement data.
The researchers use the outcomes of kinematic pose estimation strategies as enter, specializing in single-person dynamic motions from dance, strolling and sports activities. These inputs can produce correct general poses however wrestle with contacts and dynamics. A physics-based trajectory optimization subsequently enforces dynamics on the enter movement, and the researchers leverage a reduced-dimensional physique mannequin with centroidal dynamics and contact constraints to supply physically-valid motions that intently match the inputs.
As in earlier work, to get better full-body movement the researchers assume there is no such thing as a digital camera movement and the total physique is seen. This permits the strategy to realize extremely dynamic motions with out sacrificing bodily accuracy.
The researchers performed in depth qualitative and quantitative evaluations of the contact estimation and movement optimization strategies. It was proven that the proposed technique considerably enhances the realism of inferred motions over state-of-the-art strategies, and additionally estimates varied bodily properties that may be helpful for future inference of scene properties and motion recognition.
This workforce additionally identifies some analysis limitations. For instance, video optimization may be very costly, and the bodily optimization course of can take from half-hour to at least one hour for only a two-second (69 frames) video clip. Researchers hope to discover a extra environment friendly implementation technique to hurry up execution on this regard.
The paper Contact and Human Dynamics from Monocular Video is on arXiv. Click here to go to the mission web page.
Analyst: Yuqing Li | Editor: Michael Sarazen; Yuan Yuan
Synced Report | A Survey of China’s Artificial Intelligence Solutions in Response to the COVID-19 Pandemic — 87 Case Studies from 700+ AI Vendors
This report presents a have a look at how the Chinese authorities and enterprise homeowners have leveraged synthetic intelligence applied sciences within the battle in opposition to COVID-19. It can be accessible on Amazon Kindle.
Click here to seek out extra studies from us.
We know you don’t need to miss any story. Subscribe to our common Synced Global AI Weekly to get weekly AI updates.
