*********************************
There is now a CONTENT FREEZE for Mercury while we switch to a new platform. It began on Friday, March 10 at 6pm and will end on Wednesday, March 15 at noon. No new content can be created during this time, but all material in the system as of the beginning of the freeze will be migrated to the new platform, including users and groups. Functionally the new site is identical to the old one. webteam@gatech.edu
*********************************
Title : Robust approaches and optimization for 3D data
Rahul Sawhney
Robotics Ph.D. Candidate
School of Interactive Computing
Georgia Institute of Technology
Date: Wednesday, November 29, 2017
Time: 3pm - 5pm EST
Location: KACB 1315
Committee:
Dr. Charles Isbell (Advisor, School of Interactive Computing, Georgia Institute of Technology)
Dr. Henrik Christensen (Department of Computer Science and Engineering, University of California, San Diego)
Dr. Byron Boots (School of Interactive Computing, Georgia Institute of Technology)
Dr. Patricio Vela (School of Electrical and Computer Engineering, Georgia Institute of Technology)
Dr. Fuxin Li (School of Electrical Engineering and Computer Science, Oregon State University)
Abstract :
Real world data is always imperfect, and frequently corrupt. It has noise - an external, unknown, variability which cannot be predicted. For a procedure to be utile in the physical world, it often needs to be able to properly accommodate / tackle this unpredictability - it needs to be robust. This thesis discusses some approaches at representation and optimization levels for robust analysis of 3D data.
We consider robust representations for association and analysis tasks involving multiple views and scenes, in structural and indoor type settings. The importance of leveraging macro scale 3D geometry is elucidated. We discuss how this results in characterizations that are highly discriminative, and robust to noise, viewpoint changes, occlusions and related challenges. This is exemplified through their performance in settings that are significantly less restrictive than ones under which existing methods operate. Under high noise and variability, the purely geometric approaches also compare well with solutions based on RGB / RGB-D and ones based on CNNs.
We then look at robustness at a more intrinsic level - through methods for robust estimation and optimization. A robust loss function with desirable combination of properties for exact estimation and outlier suppression is introduced. And a related methodology for effective optimization of nonsmooth, nonconvex objectives is proposed. A simple to apply nonparametric, mean shift based, approach for robust mode seeking is discussed as well.
