Research Teaches AI How People Move with Internet Videos

New ideas continuing to emerge.   Here in how video can be enhanced, utilized. 

Research Teaches AI How People Move with Internet VideosThe Michigan Engineer News Center

The project enables neural networks to model how people are positioned based on only partial views of their bodies, like perspective shots in instructional videos or vlogs

Research at the University of Michigan (UM) has led to a breakthrough in video training for neural network models, enabling the models to simulate people's movement based on partial views of their bodies. UM's David Fouhey and Chris Rockwell first cropped the networks' earlier training dataset to more closely resemble online videos; they then retrained existing models on these cropped videos, facilitating more reasonable results with new data from Internet videos. The retrained model yielded significantly improved approximations of people's positions compared to two human three-dimensional mesh recovery methods. A second technique enabled models to self-train on unlabeled videos, in order to make good guesses without knowing the solution. Rockwell said, "Modeling people is a step towards understanding them, and before this it was really tough to understand people in consumer videos. With these techniques we can much more readily recognize them.