Overlapping camera clustering through dominant sets for scalable 3D reconstruction

Scalability is a great issue in modern large-scale 3D reconstruction systems.
Most Structure-from-Motion (SFM) and Multi-View-Stereo (MVS) algorithms work by
considering all images at once.

This work provides a novel camera view clustering method which works on SfM data and produces
a set of overlapping clusters suited to be processed independently by the consequent MVS.
Overlaps are important to avoid the creation of holes between clusters. Our method elegantly
integrates the clustering and overlap handling. This results in much cleaner, compacter clusters
which in the consequent MVS steps results in more detailed 3D models.


View Clustering!
Overlapping camera clustering through dominant sets for scalable 3D reconstruction
M. Mauro, H. Riemenschneider, L. Van Gool, R. Leonardi, BMVC 2013 (PDF)


The results of our method is a set of clusters of images here indicated by colors dots.
The gray squares are the overlapping cameras present in multiple clusters to ensure good coverage.

CMVS vs. our clustering for Notre Dame, Paris dataset - our clusters are more compact

CMVS vs. our clustering for Fraumunster, Zurich dataset - our overlaps are better placed

CMVS vs. our clustering for Hall dataset - our overlaps are evenly spread out


Fraumunster dataset (98 images)
This dataset contains 98 undistorted images together with the bundler.out and list.txt file.
Fraumunster is an old monastery church from 853 known for its famous Chagall windows.
The photos in this set are taken by Till Kroeger and are free to use for research purposes.

This package contains the code for running the dominant set clustering.
Input is the set of images, bundler.out and list.txt files and
optional parameters for max/min size of clusters and overlaps.
See README.txt for details and licensing.

This page has been edited by Hayko Riemenschneider