Structure from Motion (SfM)

from K. Mankoff …A newish class of algorithms that have been classified under the name “structure from motion” or SfM. While these algorithms have been in use in the robotics / computer science / other domain for over a decade (Microsoft released an implementation titled “PhotoSynth” several years ago), they have not yet been widely used in the Earth Science domain. This is changing.

Three papers have come out in the past few months introducing SfM to Earth Scientists, providing case studies:

from Adam Steer …

1. It is possible to make 3D models using photographs using freely
available software (eg SfM toolkit, bundler, CMVS, PMVS), but packages
like Photoscan (mentioned by Matt) are vastly easier to use and
contain already the tools required to geolocate your point cloud, and
do some error analysis. I’ve built both an implementation of
bundler/CMVS/PMVS in Ubuntu Linux, attempted to install and run
SfMtoolkit in Windows 7, and managed to get my hands on Photoscan for
Linux. I don’t use bundler-> CMVS or SfMtoolkit anymore. One key
reason is that I can give photoscan apriori image coordinates for use
in its bundle adjustment, and get some ideas about predicted vs
adjusted camera positions (error). The other is that I don’t have to
chase *nix libraries for every new build.

2. On the down side, SfM algorithms struggle a little on ice, since
they rely on inhomogeneous imagery to work with. It is very easy to
make models of say, Antarctic stations. It is not very easy to make
great models of sea ice unless there are lots of features and the
light is good. Mountain glaciers or snowpacks on sunny days, however,
should provide great subjects!