Towards Large-scale Multi-view Stereo Reconstruction

PhD Qualifying Examination


Title: "Towards Large-scale Multi-view Stereo Reconstruction"

by

Mr. Siyu ZHU


Abstract:

The emerging techniques handling large-scale 3D reconstruction suggest the 
possibility of reconstructing city-scale scenes from images on the 
Internet or super sized dataset and we can even build Rome in a day using 
millions of photographs in a parallel environment.

The key ingredients towards this goal are almost in place. Structure from 
motion methods use feature correspondences from matched image pairs to 
estimate precise camera geometry, Multi-View Stereo algorithms base on the 
camera geometry to propagate dense 3D points, and post-processing 
approaches take dense 3D points as inputs and produce mesh models. 
However, due to the astronomical reconstruction problem size, different 
and complicated image capturing conditions, and the increasing requirement 
for efficiency, problems gradually surface when directly applying current 
Structure from Motion and Multi-View Stereo to large-scale unstructured 
reconstruction pipeline.

In this survey, the state of the art reconstruction engine Quasi-Dense is 
utilised as an example, and the emerging approaches tacking the 
large-scale reconstruction problems are discussed under the Quasi-Dense 
framework. First, the divide-and-conquer, re-sampling, and parallelism 
techniques handling large-scale Structure from Motion are introduced. 
Then, we present the keys to distributed and parallel Multi-View Stereo, 
which propagate dense points and are the essence to the final 
reconstruction results. Finally, the important elements of the mesh 
generation process are briefly introduced.


Date:			Wednesday, 11 June 2014

Time:                  	2:00pm - 4:00pm

Venue:                  Room 3501
                         Lifts 25/26

Committee Members:	Prof. Long Quan (Supervisor)
  			Prof. Chi-Keung Tang (Chairperson)
  			Dr. Huamin Qu
  			Prof. Chiew-Lan Tai


**** ALL are Welcome ****