3D Reconstruction of Urban Scenes from Street-side and Airborne Imagery

András Bódis-Szomorú

Ph.D. Dissertation ETH Zurich 2018
Defended in Dec 2017
Accepted on the recommendation of
Prof. Dr. Luc Van Gool, Prof. Dr. Konrad Schindler and Dr. Florent Lafarge


3D models produced from this work have received media coverageYou can read about the following 3D computer vision / geometry processing topics in my thesis.


Real-time 3D point cloud reconstruction from video at 25 fps on CPU for mobile mapping.


Hierarchical Unregularized Superpixel (HUNS) algorithm outperforming methods like SLIC, ERS or SEEDS.


Image-based modelling by combining sparse 3D points and multi-view plane segmentation over superpixels.


Compact surface mesh reconstruction from images via sparse 3D points and vertex depth optimization.


Efficient conversion of dense height or depth maps into a compact mesh  via fast piecewise-planar partitioning.


Efficient surface reconstruction fusing street-side and point clouds via ray shooting into a tetrahedralization.


Segmentation of airborne 3D models into buildings, vegetation, ground, water in spite of geometric defects.


Fast multi-view texturing of urban surface models jointly from airborne and street-side data.


Image-based facade mesh segmentation into semantic parts and ways to reduce computational burden.



Defense Slides

Project Video
(This work contributed methods demonstrated from around minute 2:00 to 10:00 in the video)

Play Video


author = {András Bódis-Szomorú},
title = {3D Reconstruction of Urban Scenes from Street-Side and Airborne Imagery},
publisher = {ETH Zurich},
school = {ETH Zurich}
year = {2018},
language = {en},
size = {262 p.},
DOI = {10.3929/ethz-b-000286410},
keywords = {Multi-view reconstruction; Structure-from-Motion; Computer Vision; Superpixels; Surface Mesh; Image-based modelling; Semantic Segmentation; Urban modelling; Volumetric reconstruction; Multi-view texturing; Multi-view stereo; Image Processing; Airborne photogrammetry; Mobile mapping},