Stuart Anderson, PhD (Meta)
Topic: Quality Measurement in Metric Telepresence: Sampling and Ablation Policies for Scaling Codec Avatars
The mission of the Codec Avatars Lab within Meta Reality Labs Research is to develop technologies that deliver immersive social telepresence that’s indistinguishable from reality to anyone on earth. One of these technologies is the Universal Avatar – a model of human appearance over a broad population that can be used to quickly generate personal avatars for specific individuals. Training Universal Avatar models requires large datasets of high quality images of humans that disentangle identity, expression and environmental conditions. To collect these datasets the lab builds camera domes that can produce several petabytes of data per hour. To make storing and processing this data feasible the lab has developed data ablation policies to reduce the total volume of data without impacting the quality of models trained from that data. This talk will discuss the process of developing, validating and deploying data ablation policies on large datasets.
Stuart Anderson leads the Instant Codec Avatar group at Meta Reality Labs, researching methods that enable delivery of metric telepresence experiences at internet scale. Previously he led robotics research engineering at Meta’s Fundamental AI Research Lab (FAIR) and founded Nvidia’s Pittsburgh robotics program. Stuart received his PhD in robotics from Carnegie Mellon University in 2013. His research interests include embodied learning, with a focus on dexterous and agile behaviors.