GTC Silicon Valley-2019: Unsupervised Learning of Depth, Odometry, Flow and Segmentation using Competitive Collaboration
Note: This video may require joining the NVIDIA Developer Program or login
GTC Silicon Valley-2019 ID:S9575:Unsupervised Learning of Depth, Odometry, Flow and Segmentation using Competitive Collaboration
Anurag Ranjan(Max Planck Institute for Intelligent Systems),Jonas Wulff(MIT)
Learn about Competitive Collaboration, a framework that facilitates joint learning among several neural networks by introducing competition and collaboration. Competitive Collaboration is a three-player game in which two adversaries compete for a resource that is regulated by a moderator, where the moderator trains by a consensus between the adversaries. We'll describe how we apply our framework for joint unsupervised learning to four problems in computer vision single-image depth prediction, camera motion estimation, optical flow, and motion segmentation. These problems are coupled by geometry of the world and so geometric constraints are exploited to facilitate learning without the need for labels. We will show that joint learning using our framework achieves state-of-the art results on all the subproblems among unsupervised methods.