Towards Large Scale 3D Reconstruction

PhD Thesis Proposal Defence


Title: "Towards Large Scale 3D Reconstruction"

by

Mr. Runze ZHANG


Abstract:

Due to the popularization of smart-phones, consumer drones and social 
networks, we can obtain very large scale high resolution images. Rather 
than two- dimensional images, users prefer to record, browse and reprint 
their favorite objects or scenes with three-dimensional models. The 3D 
model requirements and large scale input available images stimulate the 
industry to provide more accurate and scalable 3D reconstruction 
techniques to use those images to recover 3D models of scenes and objects 
required by users.

A complete 3D reconstruction system contains two main key steps to recover 
3D models from images, namely Structure-from-Motion and Multiple View 
Stereo. Structure-from-Motion recovers camera poses of each image and 
sparse point positions, and Multiple View Stereo recovers 3D 
representation of scenes or objects in the images. A Structure-from-Motion 
pipeline is constructed by feature detection, image matching, camera 
registration and global bundle adjustment. Currently, feature detection, 
image matching and camera registration for very large scale image 
data-sets have been realized in a distributed manner. However the global 
bundle adjustment, which optimizes camera poses and will influence the 3D 
model quality, can be still implemented in one machine. The previous large 
scale Structure-from-Motion methods have to ignore or simplify the global 
bundle adjustment because of the memory limitation of one machine, which 
will finally affect the 3D model quality. In this proposal, we propose a 
distributed method based on space division to accomplish the global bundle 
adjustment, so that the whole Structure-from-Motion pipeline can be 
implemented distributedly. A Multiple View Stereo pipeline includes dense 
reconstruction and surface reconstruction. However, dense reconstruction 
algorithms have to load all required images into the memory, which is 
impossible for large scale image data-sets. Therefore, how to select 
suitable images and cluster images to divide the large dense 
reconstruction problem are important for the quality and scalability of 
dense reconstruction algorithms. In this proposal, similar with the 
distributed method for global bundle adjustment, we propose a space 
division based method to select and cluster images to obtain high quality 
dense point clouds in the dense reconstruction process, so that the whole 
3D reconstruction pipeline from image input to dense point clouds can be 
implemented in a distributed manner. Between Structure-from-Motion and 
Multiple View Stereo, if positional measurements from other sources are 
available, we can try to utilize them to improve the accuracy of camera 
poses. In this proposal, we propose a new positional measurement fusion 
method as the application of the proposed large scale optimization method 
to improve the result of Structure-from-Motion.

Combined with the proposed large scale optimization method in 
Structure-from-Motion and image selection and clustering method for dense 
reconstruction, our 3D reconstruction system can deal with large scale 
image data-sets in a whole distributed manner to produce high quality 3D 
models automatically and efficiently.


Date:			Thursday, 29 March 2018

Time:                  	10:00am - 12:00noon

Venue:                  Room 3494
                         (lifts 25/26)

Committee Members:	Prof. Long Quan (Supervisor)
  			Prof. Chiew-Lan Tai (Chairperson)
 			Prof. Huamin Qu
 			Dr. Pedro Sander


**** ALL are Welcome ****