3. How ASP works

The Stereo Pipeline package contains command-line and GUI programs that convert a stereo pair consisting of images and cameras into a 3D “point cloud” image. This is an intermediate format that can be passed along to one of several programs that convert a point cloud into a mesh for 3D viewing, a gridded digital terrain model (DTM) for GIS purposes, or a LAS/LAZ point cloud.

There are a number of ways to fine-tune parameters and analyze the results, but ultimately this software suite takes images and builds models in a mostly automatic way.

To create a point cloud file, not necessarily of best quality, for now, one simply passes two image files to the parallel_stereo (Section 16.47) command:

parallel_stereo --stereo-algorithm asp_bm \
  left_image.cub right_image.cub results/run

Here it is assumed that the PATH and ISISDATA environmental variables have been set, as shown in Section 2.

The .cub file format is used for non-Earth images (Section 4.2). For Earth, the images are usually in the .tif format (Section 5).

Higher quality results, at the expense of more computation, can be achieved by running:

parallel_stereo --alignment-method affineepipolar \
  --stereo-algorithm asp_mgm --subpixel-mode 3    \
  left_image.cub right_image.cub results/run

The option --subpixel-mode 9 is faster and still creates decent results (with asp_mgm/asp_sgm). The best quality will likely be obtained with --subpixel-mode 2, but this is even more computationally expensive.

It is very recommended to read Section 6, which describes other alignment methods and stereo algorithms.

The above commands will decompose the images in tiles to run in parallel, potentially on multiple machines (Section 16.47).

Or the stereo_gui frontend can be invoked, with the same options, as described in Section 16.64. This tool makes it possible to manually select smaller clips on which to run parallel_stereo.

The string results/run is an arbitrary output prefix. All parallel_stereo output files will be in the results directory and start with output. See Section 19 for the list of output files.

A a visualizable mesh or a DTM file can be made with the following commands (the results/run-PC.tif and results/run-L.tif files are created by the parallel_stereo program above):

point2mesh results/run-PC.tif results/run-L.tif
point2dem results/run-PC.tif

Here a custom projection may need to be used closer to poles (Section 16.52).

Visualization is further discussed in Section 6.2.

A produced DEM may need to be aligned to some pre-existing reference (Section 16.49).

If the positions and orientations of the cameras are not known well then bundle adjustment may be necessary (Section 12).

What follows are two examples of processing non-Earth data. An example using Earth data is in Section 5. The various stereo algorithms are discussed in Section 6. More examples can be found in Section 8.

4. Tutorial: Processing planetary data (non-Earth)

4.1. Lightning-fast example using Lunar images

This example is designed to have the user create useful results with ASP using Lunar data 10 minutes or less. It does not require a download of ISIS or ISIS data (which can be a couple of hundreds of GB) because it uses the CSM camera model (Section 8.12). The steps to process it are as follows:

• Get ASP per the installation page (Section 2).

• Fetch and extract the example dataset as:

  wget https://github.com/NeoGeographyToolkit/StereoPipelineSolvedExamples/releases/download/LRONAC/LRONAC_example.tar
  tar xfv LRONAC_example.tar
  cd LRONAC_example
  

• Start stereo_gui (Section 16.64) with a selection of clips:

  stereo_gui M181058717LE_crop.cub M181073012LE_crop.cub \
    M181058717LE.json M181073012LE.json                  \
    --alignment-method local_epipolar                    \
    --left-image-crop-win 2259 1196 900 973              \
    --right-image-crop-win 2432 1423 1173 1218           \
    --stereo-algorithm asp_mgm --subpixel-mode 9         \
    run/run
  

The crop windows from above will show up as red rectangles.

Choose from the menu Run -> Run parallel_stereo. When finished, quit the GUI and run from the command line:

point2dem --errorimage run/run-PC.tif --orthoimage run/run-L.tif

Open the computed DEM and orthoimage as:

stereo_gui run/run-DEM.tif run/run-DRG.tif

Right-click on the DEM on the left and choose to toggle hillshading to show the DEM hillshaded. See the figure below for the output.

How to get higher quality results is described in Section 6.

For other examples, see Section 8.

Fig. 4.1 Example of produced DEM and orthoimage using LRO NAC stereo pair M181058717LE and M181073012LE and CSM cameras. How to obtain and prepare the inputs is discussed in Section 8.8.

4.2. Example using Mars MOC images

The data set that is used in the tutorial and examples below is a pair of Mars Orbital Camera (MOC) [MalinDanielsonIngersoll+92, MalinEdgett01] images whose PDS Product IDs are M01/00115 and E02/01461. This data can be downloaded from the PDS directly, or they can be found in the examples/MOC directory of your Stereo Pipeline distribution.

These raw PDS images (M0100115.imq and E0201461.imq) need to be converted to .cub files and radiometrically calibrated. You will need to be in an ISIS environment (Section 2.1.1), usually via a conda activate command which sets the ISISROOT and ISISDATA environment variables; we will denote this state with the ISIS> prompt.

Then you can use the mocproc program, as follows:

ISIS> mocproc from=M0100115.imq to=M0100115.cub Mapping=NO
ISIS> mocproc from=E0201461.imq to=E0201461.cub Mapping=NO

There are also Ingestion and Calibration parameters whose defaults are YES which will bring the image into the ISIS format and perform radiometric calibration. By setting the Mapping parameter to NO, the resultant file will be an ISIS cube file that is calibrated, but not map-projected. Note that while we have not explicitly run spiceinit, the Ingestion portion of mocproc quietly ran spiceinit for you (you’ll find the record of it in the ISIS Session Log, usually written out to a file named print.prt). Fig. 4.2 shows the results at this stage of processing.

Fig. 4.2 This figure shows E0201461.cub and M0100115.cub open in ISIS’s qview program. The view on the left shows their full extents at the same zoom level, showing how they have different ground scales. The view on the right shows both images zoomed in on the same feature.

See Section 8 for many solved examples, including how to preprocess the data with tools specific for each mission.

Once the .cub files are obtained, it is possible to run parallel_stereo right away:

ISIS> parallel_stereo E0201461.cub M0100115.cub    \
        --alignment-method affineepipolar          \
        -s stereo.default.example results/output

In this case, the first thing parallel_stereo does is to internally align (or rectify) the images, which helps with finding stereo matches. Here we have used affineepipolar alignment. Other alignment methods are described in Section 6.1.2.

If your data has steep slopes, mapprojection can improve the results. See Section 6.1.7 and Section 11.4.3.

See Section 6 for a more in-depth discussion of stereo algorithms.

5. Tutorial: Processing Earth DigitalGlobe/Maxar images

In this chapter we will focus on how to process Earth images, or more specifically DigitalGlobe/Maxar data. This example is different from the one in the previous chapter in that at no point will we be using ISIS utilities. This is because ISIS only supports NASA instruments, while most Earth images comes from commercial providers.

In addition to DigitalGlobe/Maxar’s satellites, ASP supports any Earth images that uses the RPC camera model format. How to process such data is described in Section 8.19, although following this tutorial may still be insightful even if your data is not from DigitalGlobe/Maxar.

If this is your first time running ASP, it may be easier to start with ASTER data (Section 8.17), as its images are free and much smaller than DigitalGlobe’s. A ready-made example having all inputs, outputs, and commands, is provided there.

DigitalGlobe provides images from QuickBird and the three WorldView satellites. These are the hardest images to process with Ames Stereo Pipeline because they are exceedingly large, much larger than HiRISE images (the GUI interface can be used to run stereo on just a portion of the images). There is also a wide range of terrain challenges and atmospheric effects that can confuse ASP. Trees are particularly difficult for us since their texture is nearly nadir and perpendicular to our line of sight. It is important to know that the driving force behind our support for DigitalGlobe/Maxar images is to create models of ice and bare rock. Those are the type of images that we have tested with and have focused on. If we can make models of wooded or urban areas, that is a bonus, but we can’t provide any advice for how to perform or improve the results if you choose to use ASP in that way.

The camera information for DigitalGlobe/Maxar images is contained in an XML file for each image. In addition to the exact linear camera model, the XML file also has its RPC approximation. In this chapter we will focus only on processing data using the linear camera model. For more detail on RPC camera models we refer as before to Section 8.19.

Our implementation of the Digital Globe linear camera model accounts for the sensor geometry, velocity aberration and atmospheric refraction (Section 11.4.2). These corrections will shift point locations by over a meter for some images. However this is still smaller error than the error from measurement of the spacecraft’s position and orientation. The latter can be corrected using bundle adjustment, ideally used with ground control points (Section 16.5). Alternatively, the pc_align tool discussed in Section 6.1.12 can be used to align the terrain obtained from ASP to an accurate set of ground measurements.

In the next two sections we will show how to process unmodified and map-projected variants of WorldView images. The images we are using is from the free stereo pair labeled “System-Ready (1B) Stereo, 50cm” which captures the city of Stockholm, found on DigitalGlobe/Maxar’s website (https://www.digitalglobe.com/samples). These images represent a non-ideal problem for us since this is an urban location, but at least you should be able to download these images yourself and follow along.

5.1. Supported products

ASP can only process Level 1B satellite images, and cannot process DigitalGlobe’s aerial images or orthorectified images (see the product info).

5.2. Processing raw

After you have downloaded the example stereo images of Stockholm, you will find a directory titled:

056082198020_01_P001_PAN

It has a lot of files and many of them contain redundant information just displayed in different formats. We are interested only in the TIF or NTF images and the similarly named XML files.

Some Worldview folders will contain multiple image files. This is because DigitalGlobe/Maxar breaks down a single observation into multiple files for what we assume are size reasons. These files have a pattern string of “_R[N]C1-”, where N increments for every subframe of the full observation. The tool named dg_mosaic can be used to mosaic (and optionally reduce the resolution of) such a set of sub-observations into a single image file and create an appropriate camera file:

dg_mosaic 12FEB16101327*TIF --output-prefix 12FEB16101327

and analogously for the second set. See Section 16.20 for more details. The parallel_stereo program can use either the original or the mosaicked images. This sample data only contains two image files so we do not need to use the dg_mosaic tool.

Since we are ingesting these images raw, it is strongly recommended that you use affine epipolar alignment to reduce the search range. The parallel_stereo command and a rendering of the results are shown below.

parallel_stereo -t dg --stereo-algorithm asp_mgm      \
  --subpixel-mode 9 --alignment-method affineepipolar \
  12FEB16101327.r50.tif 12FEB16101426.r50.tif         \
  12FEB16101327.r50.xml 12FEB16101426.r50.xml         \
  run/run

As discussed in Section 3, one can experiment with various tradeoffs of quality versus run time by using various stereo algorithms, and use stereo in parallel or from a GUI. For more details, see Section 6.

How to create a DEM and visualize the results of stereo is described in Section 6.2.

Consider using above the options --enable-correct-velocity-aberration and --enable-correct-atmospheric-refraction to improve the positioning of the produced DEMs, unless using bundle adjustment (Section 17). In either case, pc_align (Section 16.49) can be used to align the produced DEM to a desired reference terrain.

Fig. 5.1 Example WorldView image section and colorized height map.

It is important to note that we could have performed stereo using the approximate RPC model instead of the exact linear camera model (both models are in the same XML file), by switching the session in the parallel_stereo command above from -t dg to -t rpc. The RPC model is somewhat less accurate, so the results will not be the same, in our experiments we’ve seen differences in the 3D terrains using the two approaches of 5 meters or more.

Many more stereo processing examples can be found in Section 8.

5.3. Processing map-projected images

ASP computes the highest quality 3D terrain if used with images map-projected onto a low-resolution DEM that is used as an initial guess. This process is described in Section 6.1.7.

5.4. Dealing with clouds

Clouds can result in unreasonably large disparity search ranges and a long run-time. It is then suggested to mapproject the images (Section 6.1.7).

With our without mapprojection, one can reduce the computed search range via --max-disp-spread (Section 17). Use this with care. Without mapprojection and with steep terrain, the true spread of the disparity can, in rare cases, reach a few thousand pixels. This is best used with mapprojected images, when it is likely to be under 150-200, or even under 100.

If a reasonable DEM of the area of interest exists, the option --ip-filter-using-dem can be used to filter out interest points whose heights differ by more than a given value than what is provided by that DEM. This should reduce the search range. Without a DEM, the option --elevation-limit can be used and should have a similar effect.

Another option (which can be used in conjunction with the earlier suggestions) is to tighten the outlier filtering in the low-resolution disparity D_sub.tif (Section 19), for example, by setting --outlier-removal-params 70 2 from the default 95 3 (Section 17). Note that decreasing these a lot may also filter out valid steep terrain.

If a run failed because of a large disparity search range, D_sub.tif should be deleted, parameters adjusted as above, and one should run stereo_corr with the same arguments that parallel_stereo was run before (except those used for tiling and number of processes, etc.), while adding the option --compute-low-res-disparity-only. Then examine the re-created D_sub.tif with disparitydebug (Section 16.21) and the various search ranges printed on screen.

When D_sub.tif is found to be reasonable, parallel_stereo should be re-run with the option --resume-at-corr.

See also Section 6.1.10 which offers further suggestions for how to deal with long run-times.

5.5. Handling CCD boundary artifacts

DigitalGlobe/Maxar WorldView images [Glo] may exhibit slight subpixel artifacts which manifest themselves as discontinuities in the 3D terrain obtained using ASP. We provide a tool named wv_correct, that can largely correct such artifacts for World View-1 and WorldView-2 images for most TDI.

Note that Maxar (DigitalGlobe) WorldView-2 images with a processing date (not acquisition date) of May 26, 2022 or newer have much-reduced CCD artifacts, and for those this tool will in fact make the solution worse, not better. This does not apply to WorldView-1, 3, or GeoEye-1.

This tool can be invoked as follows:

wv_correct image_in.ntf image.xml image_out.tif

The corrected images can be used just as the originals, and the camera models do not change. When working with such images, we recommend that CCD artifact correction happen first, on original un-projected images. Afterward images can be mosaicked with dg_mosaic, map-projected, and the resulting data used to run stereo and create terrain models.

This tool is described in Section 16.69, and an example of using it is in Fig. 5.2.

Fig. 5.2 Example of a hill-shaded terrain obtained using stereo without (left) and with (right) CCD boundary artifact corrections applied using wv_correct.

Another source of artifacts in linescan cameras, such as from DigitalGlobe, is jitter. ASP can solve for it using a jitter solver (Section 16.35).

5.6. Dealing with terrain lacking large-scale features

Stereo Pipeline’s approach to performing correlation is a two-step pyramid algorithm, in which low-resolution versions of the input images are created, the disparity map (output_prefix-D_sub.tif) is found, and then this disparity map is refined using increasingly higher-resolution versions of the input images (Section 11.2).

This approach usually works quite well for rocky terrain but may fail for snowy landscapes, whose only features may be small-scale grooves or ridges sculpted by wind (so-called zastrugi) that disappear at low resolution.

A first attempt at solving this is to run parallel_stereo with:

--corr-seed-mode 0 --corr-max-levels 2

This will prevent creating a low-resolution disparity which may be inaccurate in this case. (Note that interest points which are computed before this are found at full resolution, so they should turn out well.) Here, ASP will run correlation with two levels, so the lower initial resolution is a factor of 4 coarser than the original, which will hopefully prevent small features from being lost.

If that is not sufficient or perhaps not fast enough, Stereo Pipeline provides a tool named sparse_disp to create the low-resolution initial disparity output_prefix-D_sub.tif based on full-resolution images, yet only at a sparse set of pixels for reasons, of speed. This low-resolution disparity is then refined as earlier using a pyramid approach, but again with fewer levels.

Fig. 5.3 Example of a difficult terrain obtained without (left) and with (right) sparse_disp. (In these DEMs there is very little elevation change, hence the flat appearance.)

This mode can be invoked by passing to parallel_stereo the option --corr-seed-mode 3. Also, during pyramid correlation it is suggested to use somewhat fewer levels than the default --corr-max-levels 5, to again not subsample the images too much and lose the features.

Here is an example:

parallel_stereo -t dg --corr-seed-mode 3            \
  --corr-max-levels 2                               \
  left_mapped.tif right_mapped.tif                  \
  12FEB12053305-P1BS_R2C1-052783824050_01_P001.XML  \
  12FEB12053341-P1BS_R2C1-052783824050_01_P001.XML  \
  dg/dg srtm_53_07.tif

If sparse_disp is not working well for your images you may be able to improve its results by experimenting with the set of sparse_disp options which can be passed into parallel_stereo through the --sparse-disp-options parameter. sparse_disp has so far only been tested with affineepipolar image alignment so you may not get good results with other alignment methods.

The sparse_disp tool is written in Python, and it depends on a version of GDAL that is newer than what we support in ASP and on other Python modules that we don’t ship. It is suggested to to use the Conda Python management system at

https://docs.conda.io/en/latest/miniconda.html

to install these dependencies. This can be done as follows:

conda create --name sparse_disp -c conda-forge python=3.9 gdal
conda activate sparse_disp
conda install -c conda-forge scipy pyfftw

Assuming that you used the default installation path for conda, which is $HOME/miniconda3, before running the parallel_stereo command, as shown above, one needs to set:

export ASP_PYTHON_MODULES_PATH=$HOME/miniconda3/envs/sparse_disp/lib/python3.9/site-packages

It is very important that the same version of Python be used here as the one shipped with ASP. Note that if GDAL is fetched from a different repository than conda-forge, one may run into issues with dependencies not being correct, and then it will fail at runtime.

5.7. Processing multi-spectral images

In addition to panchromatic (grayscale) images, the DigitalGlobe/Maxar satellites also produce lower-resolution multi-spectral (multi-band) images. Stereo Pipeline is designed to process single-band images only. If invoked on multi-spectral data, it will quietly process the first band and ignore the rest. To use one of the other bands it can be singled out by invoking dg_mosaic (Section 5.2) with the --band <num> option. We have evaluated ASP with DigitalGlobe/Maxar’s multi-spectral images, but support for it is still experimental. We recommend using the panchromatic images whenever possible.