Problem with finding interest points

[cecilia martinelli]

Hello,

I am working with ADS images and trying to correlate them using parallel_stereo. The problem is that ASP can't find any interest points between the images (which is strange). As a workaround, I tried adding the --skip-rough-homography and --corr-search options. However, ASP still stops as soon as it fails to find any interest points.

Why does this could happen? And is there a way to force it to continue despite not finding IPs?

Any insights or recommendations would be greatly appreciated!

Thanks

Cecilia

my whole command line is:

 parallel_stereo --correlator-mode pre_crop.tif post_crop.tif $mgm/corr --alignment-method none --corr-timeout 500 --stereo-algorithm asp_final_mgm --corr-kernel 9 9 --cost-mode 3 --subpixel-mode 10 --threads-singleprocess $threads --threads-multiprocess $mpthread --processes $mpproc --xcorr-threshold -1 --min-xcorr-level 1 --corr-seed-mode 0 --sgm-collar-size 320  --min-num-ip 10 --ip-per-tile 500 --skip-rough-homography --corr-search -15 -15 15 15  --rm-cleanup-passes 1 --filter-mode 2 --rm-threshold 4.0 --rm-min-matches 40 --rm-half-kernel 9 9

[Oleg Alexandrov]

Interest point matching fails if the images are too different, usually. It is suggested to examine your inputs. It may also help to see what messages get printed before it fails.

> And is there a way to force it to continue despite not finding IPs?

With the latest build, and with the option  --corr-search, the program will continue without interest points. The latest build can be found here: https://github.com/NeoGeographyToolkit/StereoPipeline/releases.

--
You received this message because you are subscribed to the Google Groups "Ames Stereo Pipeline Support" group.
To unsubscribe from this group and stop receiving emails from it, send an email to ames-stereo-pipeline...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/ames-stereo-pipeline-support/fffc6a26-60df-43b5-9d51-5d2ad26ddcf9n%40googlegroups.com.

[cecilia martinelli]

I managed to make progress with parallel_stereo by working on a different band of the image... however, now I’m facing a similar problem, again.

 I'm trying to align my ADS image to a SPOT image that covers exactly the same area, using image_align. My ADS image has a resolution of 0.4 m, while the SPOT image is 2 m, so I upsampled the SPOT image to 0.4 m. However, this time—despite switching bands again—it’s unable to find interest points between the two images, even though they appear visually similar.

The command I'm using is:

image_align SPOT_04m.tif ADS_b3.tif --alignment-transform similarity --ip-per-image 50000 -o ADS_aligned.tif

and the terminal output is as follows:

Warning: read_gdal_georeference(): WARNING! Resource file  .../SPOT_04m.tif contains a non-normal georeference.
Matching interest points between: .../SPOT_04m.tif and .../ADS_b3.tif
            Looking for IP in left image.
            Using 50 interest points per tile (1024^2 px).
            Detecting IP
            Removing IP near nodata with radius 4
            Building descriptors
            Found interest points: 0
            Looking for IP in right image.
            Using 50 interest points per tile (1024^2 px).
            Detecting IP
            Removing IP near nodata with radius 4
            Building descriptors
            Found interest points: 106538
Elapsed time in ip detection: 506.745 s.
        --> Uniqueness threshold: 0.8
           Matching: [****************************************************]
Elapsed time in ip matching: 0.000227 s.
            Matched points: 0
RANSAC Error.  Insufficient data.

ERROR: RANSAC was unable to find a fit that matched the supplied data.
Alignment transform computation failed.

About the warning, I checked my images and they are normally (north-up) georeferenced, so I'm not sure where it's coming from.

I would like to use the images as unresampled as possible, so I haven’t tried changing the georeferencing system, but I’m running out of ideas on how to move forward, so I’m open to hear any advice on the issue and how it could be resolved! 

Thanks!

Cecilia

[Alexandrov, Oleg (ARC-TI)[KBR Wyle Services, LLC]]

From the output of your command:

>Found interest points: 0

That is not good. It failed to even find anything in the image.

It depends on how upsampling was done. It is good to use bicubic interpolation for best results (with gdalwarp).

Things may work better if, at least temporality, both of these are sampled to say 1 m / pixel or 2 m / pixel, with bicubic interpolation, and see if anything changes.

The doc https://stereopipeline.readthedocs.io/en/latest/tools/image_align.html#examples suggests using parallel_stereo --correlator-mode as an initial step. That may help with more difficult data. The parallel_stereo program also has extra switches for interest points, such as modifying the interest point detection mode (--ip-detect-mode) and also has the option --individually-normalize that can help with interest points if the images have different normalizations.

If you can share privately the data I could take a closer look. 

---

From: ames-stereo-pi...@googlegroups.com <ames-stereo-pi...@googlegroups.com> on behalf of cecilia martinelli <cecilia.ma...@gmail.com>
Sent: Friday, May 16, 2025 9:01 AM
To: Ames Stereo Pipeline Support <ames-stereo-pi...@googlegroups.com>
Subject: [EXTERNAL] [BULK] Re: Problem with finding interest points

 

CAUTION: This email originated from outside of NASA.  Please take care when clicking links or opening attachments.  Use the "Report Message" button to report suspicious messages to the NASA SOC.

To view this discussion visit https://groups.google.com/d/msgid/ames-stereo-pipeline-support/d956f5c2-8788-46d3-b71a-454c3fc84e91n%40googlegroups.com.