Runze
Zhang

PhD Candidate in CSE, HKUST

Intro

What I am all about.

I am a 5th year PhD student in Department of Computer Science and Engineering, HKUST, supervised by Prof. Long Quan. Before that, I received my Bachelor of Science degree from Department of Machine Intelligence, Peking University in 2013. I am expected to graduate in June, 2018.

My research interests are the large scale 3D reconstruction and SLAM, including Structure-from-Motion, Multi-view Stereo, and other related topics. I am also one of co-founders of Altizure, which is a start-up to help users reconstruct 3D models automatically from images and provide solutions for large scale surveying and mapping.

Projects

Building the Earth in 3D

Project1
Hide

Distributed Very Large Scale Global Optimization in SfM

Recent works on large scale Structure-from-motion can tackle millions of images, but the global optimization, such as motion averaging and bundle adjustment cannot be performed on such scale data-sets due to the memory limitation of single machines. Without the global optimization, the output of 3D reconstruction may have low accuracy. Our work provides a distributed solution for very large scale global optimization problem in SfM, which has faster convergence rates and lower communication overhead. The work has been summarized in paper accepted as oral presentation by ICCV 2017.


Distributed Very Large Scale
Global Optimization in SfM


Project2
Hide

Camera Clustering for Large Scale Multi-view Stereo

Multi-view Stereo algorithms always work on stereo matching problem of two or several images. However, it is very hard to process thousands of input images for current Multi-view Stereos. Therefore, it is required to cluster input images and merge the results of different image clusters. Previous methods may generate redundant, non-uniform dense point clouds with different qualities. Our designed algorithm tries to generate dense point clouds only once with the best qualities for each image clusters, which can provide better input for the following surface reconstruction. The work has been summarized in paper accepted by ICCV 2015. This algorithm can also be applied to the image pair selection for the surface refinement problem.

Camera Clustering for
Large Scale Multi-view Stereo

Project3
Hide

Fusion of Visual SLAM and Position Measurements

Previous methods to fuse camera poses of visual SLAM and globle position measurements require registering the results of visual SLAM to the coordinate system of position measurements. If the drift of visual SLAM is too large, the visual results may be registered into wrong coordinate systmen, which leads to the difficulty of the following optimization. Besides, visual results of some degraded motion pattern cannot be registered into the poision measurement system. Hence, we propose an algorithm based on tetrahedral similarity to fuse visual and position measurement results before the coordinate system alignment. The work has been summarized in paper accepted by ACCV 2014.

Fusion of Visual SLAM and
Position Measurements

Demo

Models by
our 3D reconstruction techniques

Austin 100 km2, United States:

Publications

Contacts

Location
4204, Academic Building, HKUST
Top