Title: An Image Based Approach for 3D Reconstruction of Urban Scenes by Exploiting Architectural Symmetries

Natesh Srinivasan

Computer Science Ph.D. Student

School of Interactive Computing

College of Computing

Georgia Institute of Technology

Date: Monday, June 6th, 2016

Time: 11:00am to 1:00pm (EST)

Location: TBD

Committee:

---------------

Dr. Frank Dellaert (Advisor), School of Interactive Computing, Georgia Institute of Technology Dr. Irfan Essa (Co-Advisor), School of Interactive Computing, Georgia Institute of Technology Dr. James Hays, School of Interactive Computing, Georgia Institute of Technology Dr. Yanxi Liu, Department of Computer Science, University of Pennsylvania

Abstract:

-----------

Generation of accurate 3D models of urban environments is an important problem that needs to be addressed in the modern society and cameras offer as a cheap and easy way to achieve this. I address the problem of reconstructing urban scenes exhibiting architectural symmetries from an image or a set of images. I propose a fully automatic algorithm that can identify multiple 3D symmetries that is not restricted to translational symmetries alone but also considers rotational and bilateral symmetries in 3D. In particular I make the following three major contributions :

1. I show that a voting scheme in a polar transformation coordinates can robustly detect 3D symmetries in noisy, distorted point clouds obtained from image based reconstruction and also provide a reasonable initialization for joint optimization of 3D structure and translational and rotational symmetries.

2. In cases where it is not possible to decouple the detection of symmetry and 3D reconstruction, I show that relaxing the constraints on feature matching to generate a “clique” of matches can be a useful method for both the detection of 3D symmetries and as a solution for incorrect correspondence arising as a consequence of symmetry. A symmetry-aware RANSAC stage in SfM can extract the relevant cliques that pertain to symmetry.

3. Finally I propoe that by explicitly modelling the 3D basis element using an object-centric coordinate system, we can eliminate the manual intervention needed to generate fully automatic algorithm for detecting symmetries in 3D and also obtain a dense element-wise reconstruction.