8.1. Guidelines for selecting stereo pairs¶
When choosing image pairs to process, images that are taken with similar lighting conditions and significant surface coverage overlap are best suited for creating terrain models [BeckerArchinalHare+15]. The images should have sufficient difference in perspective, hence a reasonably large baseline, or, equivalently, a non-small convergence angle between the matching rays emanating from the two cameras, for stereo triangulation to be accurate. Yet, if the perspectives are very different, it will be challenging to compute the stereo correlation between images. A convergence angle of 10 to 60 degrees is likely reasonable.
Depending on the characteristics of the mission data set and the individual images, the degree of acceptable variation will differ. Significant differences between image characteristics increases the likelihood of stereo matching error and artifacts, and these errors will propagate through to the resulting data products.
The parallel_stereo
and bundle_adjust
programs compute the
convergence angle for input cameras. In stereo that happens at the
preprocessing stage (Section 16.52.4), with the result
printed on the screen and saved to the log file. In bundle_adjust
this computation takes place after the optimization of the cameras
finishes, and the results are saved to a file on disk
(Section 16.5.11). To find good stereo pairs, one can run
bundle adjustment on a large set of images and pick a pair with a
decent convergence angle.
Although images do not need to be mapprojected before running the
parallel_stereo
program, we recommend that you do run cam2map
(or
cam2map4stereo.py
) beforehand, especially for image pairs that
contain large topographic variation (and therefore large disparity
differences across the scene, e.g., Valles Marineris). mapprojection is
especially necessary when processing HiRISE images. This removes the
large disparity differences between HiRISE images and leaves only the
small detail for the Stereo Pipeline to compute. Remember that ISIS can
work backwards through a mapprojection when applying the camera model,
so the geometric integrity of your images will not be sacrificed if you
mapproject first.
An alternative way of mapprojection, that applies to non-ISIS images
as well, is with the mapproject
tool (Section 6.1.7).
Excessively noisy images will not correlate well, so images should be photometrically calibrated in whatever fashion suits your purposes. If there are photometric problems with the images, those photometric defects can be misinterpreted as topography.
Remember, in order for parallel_stereo
to process stereo pairs in
ISIS cube format, the images must have had SPICE data associated by
running ISIS’s spiceinit
program run on them first.