16.17. Image correlator

The parallel_stereo program (Section 16.52) can function purely as an image correlator, that is, it can find the disparity between two given images without assuming any cameras are present and without generating a point cloud.

This mode assumes that the images are already roughly aligned, up to a translation. Hence, the images can be either raw but with no large rotation/shear/scale differences, or mapprojected, or to be portions of aligned images L.tif and R.tif (Section 19) as created by ASP itself.

The produced disparity map can be used for image alignment (Section 16.33) and terrain alignment (Section 16.54.3.5).

See also corr_eval (Section 16.16) a program for estimating the quality of the disparity at each pixel.

Example:

parallel_stereo              \
  --correlator-mode          \
  --stereo-algorithm asp_mgm \
  --subpixel-mode 9          \
  left.tif right.tif         \
  run_corr/run

This will create the filtered subpixel disparity run_corr/run-F.tif (Section 19). This disparity can be preprocessed and inspected (Section 16.34.1.9).

ASP supports many stereo correlation algorithms (Section 6.1.1). It may be worth comparing the result of the asp_mgm algorithm with what is produced from regular block matching (asp_bm), especially if the images are noisy or differ in illumination.

If the expected search range is known, it can be specified with the option --corr-search (Section 17.2). Low resolution disparity computation can be skipped with --corr-seed-mode 0. These can be helpful if interest point matching or low-res disparity are problematic.

If the options --num-matches-from-disparity or --num-matches-from-disp-triplets are specified, dense matches from disparity will be produced (Section 17). These can be used in bundle adjustment (Section 12.2.4.2).

All the usual options of parallel_stereo apply. See Section 6 for a discussion regarding various quality vs speed tradeoffs. Since the images are assumed to be aligned, the program will set the alignment method to none.

To skip preprocessing (if invoked previously), or to avoid subpixel refinement or filtering, use the options --entry-point and --stop-point.