Friday, April 3
11:30 am to 12:30 pm
Newell-Simon Hall 4305
11:30 am to 12:30 pm
Newell-Simon Hall 4305
Do we need geometry for 3D, or 3D for robotics?
Abstract: In the current era of large-scale learning, a ‘bitter lesson’ has been that we should pursue general methods that scale with computation (and data) instead of ones that prioritize human knowledge of the domain. An assumed corollary of this lesson for systems pursuing 3D understanding is that explicit geometry will not be needed for these. Similarly, a sentiment in robot learning community is that intermediate representations (e.g. 3D) are not needed for learning end-to-end mappings from perception to action.
In this talk, I will first discuss earlier work from our group that supports the above theses e.g. highlighting a paradigm shift where ‘geometry-free’ neural models are being adopted for classical 3D tasks e.g. Structure-from-motion(SfM). I will then examine the role of geometry in the era of such large-scale multi-view models and present more recent works that show the value of geometric inductive biases for interpretability, accuracy, and scalability. Finally, I will present some related efforts in incorporating 3D representations and inductive biases for robot learning.