Forum : Real World & Events

Geotags, aerial synths, street synths… capturing events as they happen. If you can mark it on a map, then this is the place to talk about it.


Topic: Structure from motion

Report Abuse Report Abuse
Ross-Garland (Over 1 year ago)
Hi all,

I'm currently in the process of writing my dissertation, where I am using SfM as a means of cheap and rapid analysis of landslide monitoring. So I'll obviously need to write a section on SfM and I'm by no means computer literate at all! I'm yet to research this in depth yet, so some pointers about what to look for and what to avoid would be brilliant!

Also, if anyone could share any ideas about the application of SfM. I'm aware it has been heavily used in archaeology and has been used in avalanche monitoring etc, but I'm sure it will have much more uses.

Once again, any help would be absolutely fantastic!

Thanks, Ross Garland
Nathanael (Over 1 year ago)
Greetings, Ross, 

I'm no computer scientist or developer, but my curiosity with what made Photosynth tick has lead me to watch interviews that have gotten me familiar with some computer vision/photogrammetry terminology. 

I can't say that I'm all that familiar with your particular use case, either (nor necessarily of all of the archaeology or avalanche monitoring examples that you've found) and I haven't been forced to write anything approaching a dissertation, so it's a little difficult for me to know exactly what you're looking for in terms of information.

The first thing that came to mind when reading your query was Mark Willis' post: 
"Determining Rock Art Deterioration Through Time" ( http://palentier.blogspot.com/2011/11/determining-rock-art-deterioration.html )
Mark's Photosynth profile: http://photosynth.net/userprofilepage.aspx?user=sir_ivar
Nathanael (Over 1 year ago)
I'm not sure how far down the Computer Vision rabbit hole you'll be wanting or needing to go for your dissertation. 

It's somewhat unclear to me whether you're looking mainly for specific applications of SfM that you could use to build a use-case for landslide monitoring (what I take to be the point of your second paragraph) or whether you're asking for a jump start on some papers and videos to help your researching of SfM (possible interpretation of the first paragraph).

For applications/use cases, you might want to browse the Photogrammetry Forum: http://pgrammetry.com/
Nathanael (Over 1 year ago)
Structure from Motion is good at taking image features found by an algorithm such as David Lowe's SIFT ( http://en.wikipedia.org/wiki/Scale-invariant_feature_transform ) and estimating the 3D locations of those features as well as the poses of the cameras which took the photos which the image features are observed in. 

The output from SfM is usually limited in terms of the detail of the geometry recovered for the photographed object(s), however and either: 
1: you use the computed camera poses to help you project two photos of the same location, but taken from different positions (as in Mark's example above) onto a common 3D model to more naturally observe the changes over time (relying on the input photos for the comparison)
or 
2: you use the camera poses which SfM has supplied you to then use more processor intensive processes such as Multi-View Stereo to take smaller clusters of adjacent cameras and compute greater resolution geometry.
Nathanael (Over 1 year ago)
If you had the foresight or good fortune to take enough photos of the location beforehand (to generate a usefully detailed model of the terrain) and also enough photos after the landslide (to generate a detailed enough updated model of the terrain), then using Structure from Motion would be useful for helping you make better direct comparisons between the input photos, however with sufficient data from before and after the landslide, you could compute a far more detailed description of the change in topography by proceeding to Multi-View Stereo, manually patching any obvious errors in the resulting models (such as holes) and then aligning the before model to the after model, based on part of the terrain which has not changed, and then using a 3D modelling program to highlight where the two models diverge.
Nathanael (Over 1 year ago)
I'll point you to two informal lecture-style talks which should help you become more familiar with the lay of the land in this world and if I think of anything else that would be of use I'll post it here later. 

2009 05 22: 'Cyberspace Arriving' by Blaise Agüera y Arcas: http://video-jsoe.ucsd.edu/asx/AguerayArcas.asx
2013 02 01: 'A Trillion Photos' by Steve Seitz: http://www.youtube.com/watch?v=6MWEfpKUfRc

I just happened across this which is also userful for historical context of the photogrammetry and computer vision world: 
2011 08 15: '3D Computer Vision: Past, Present, and Future' by Steve Seitz: http://www.youtube.com/watch?v=kyIzMr917Rc

Also see the papers section at the Photogrammetry Forum for a few interesting finds, although it is by no means an organized or remotely complete list: 
http://www.pgrammetry.com/forum/viewforum.php?f=34&sid=d71882d864a415eb9f482b2ef9031cb6

Please feel free to clarify what you're looking for if I'm not hitting the mark.
You need to be Signed In to add a comment. (Are you new? Sign Up for a free account.)