Notes are from Karl's coffee talk on 11/10/16. The pdf of these notes is available on the server @ esd/share/group_seminar_presentations/coffee-talks_2016/karl_111016_sfm
Structure from Motion:
What it is: a photogrammetric technique used to estimate a third-dimensional distance from two-dimensional images (this is actually quite old). It uses “highly redundant, iterative bundle adjustment procedures” (these are pretty new) to extract common pixels from overlapping images, then solves for both the position of the camera that took each image and the relative position of the pixels with respect to the cameras. This approach is scale independent and scale must be added to the model by digitization or georeferencing.
Basics of use: Collect photographs of an object or surface at a variety of vertical and/or horizontal angles then process images to make point cloud, then TIN and can export as gridded DSM, etc. Taking the images is the most important part and model resolution is ultimately limited by the number of matched pixels, from:
1. sensor size and pixel density
2. quality of images (mostly relates to lens quality)
3. image overlap
1. “these days” any decent camera is fine. at least 5 Mpx, but more like 12Mpx is better for “professional quality"
Iphone 7 has 12 Mpx (smh),
Nikon D5 ($7k) has 20.8 Mpx,
GoPros are 8-12 Mpx
Nikon KeyMissions are pretty awesome (180-360 degree images)
2. image distortion can be a problem that is easily correctable… but reduces effective pixel resolution at image margins this is part of the reason why its important to have significant image overlap best approach: use lenses (for removable lens cameras) appropriate for the sensor size shutter-less, removable lense cameras (Sony Alphas) are pretty cool - may be better for field.
3. try for 60% overlap between images (side) avoid shiny/reflective/glittery surfaces (e.g. can use talcum powder, or cloth) in some cases, need fixed camera positions (not always realistic for field) also: can add control points inside the model and/or georeferencing of the cameras (both is best). e.g. throw a couple meter stick into a model for scale. Can get GPS for camera (but drains battery… I’ve found just holding a GPS is fine) For DEMs, need at least ~10 ground control points for a good reconstruction
There is lots of open-source options, e.g. a list on Wikipedia
but... commerical software is the best to use, still very flexible and has a gui that is easy to use. The best option is Agisoft Photoscan. Agisoft Photoscan is bulletproof, easy to use, cheap ($550/node locked licence). 30 day free trial too (full professional version) Just be aware that it’s memory intensive
Cool ideas for applications:
1. DEMs
pro: repeat scans, high-res, cheap (relatively)
con: site-specific
use drones, balloons, helikites, poles, e.g. Sensefly eBees, helikites
2. outcrops
pro: same as above, get pixel color (may correlate with lithology, etc.), issues with reflective surfaces
con: shadows could be an issue? (take lights to field?)
could be an essential compliment to field work/photography
3. fossils/artifacts/art?
pro: flexible resolution, even with lights/set-up, still cheaper than laser scanner?
con: ?
4. underwater work, e.g. flume, river and lake profiling)
? (also check out sketchfab)
Other good references:
Note that the Agisoft website has great tutotials and tutotials