Have ASP produce stereo at lower resolution than the input images #43

Open
oleg-alexandrov opened this Issue Jan 7, 2013 · 4 comments

Comments

Projects
None yet
2 participants
Owner

oleg-alexandrov commented Jan 7, 2013

It would be desirable to have stereo output a point cloud at say 2x coarser resolution than the input images. To implement that, one would need to internally sub-sample the input images and scale accordingly the cameras. The scaling of images is trivial, the scaling of the cameras could be achieved by very small modifications to the point_to_pixel(), pixel_to_vector() and camera_center() functions, regardless of camera type.

This is not urgent, I put it here to make sure we don't forget.

Contributor

SmithB commented Jan 25, 2013

For ice-sheet images, subsampling tends to blur out the texture that ASP needs to get good matches. Would it be possible instead to correlate the full-resolution images and output lower-resolution DEMs? For the sake of speed, the resolution reduction would happen before triangulation.

Owner

oleg-alexandrov commented Apr 24, 2013

This would save quite little computational time as correlation and refinement takes much longer than triangulation and DEM generation. One could as well do the remaining two steps at full res and downsample the resulting DEM.

Contributor

SmithB commented Apr 24, 2013

The benefit would be if refinement could be done only on a subset of the full-res disparities. At present, the full Bayesian refinement is often the slowest step for high-resolution pairs, so refining only a subset of the points could speed up the pipeline considerably. This is probably only important on snow surfaces where running the whole pipeline at reduced resolution tends to fail.

Owner

oleg-alexandrov commented Apr 24, 2013

I see. I'll defer the decision to Zack. In the meantime, if you have your own disparity, sparse or dense, and all you want is refinement/triangulation/DEM, it should be possible using the following workaround (which you may have figured out on your own already):

  1. Create a -D.tif file, of the same size as the input left image. All the values in it can be invalid (the third band is the valid data mask) except for the pixels you care about at which you set the disparity.
  2. Run stereo_rfne.
  3. Skip stereo_fltr, instead, copy -RD.tif to -F.tif.
  4. Run triangulation and DEM generation.

I did not try these, but I see no reason why it should not work.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment