Using VPI stereo_disparity with cuda backend yeids errorenous disparity images at high resolution

[not_a_bot]

I am able to generate disparity maps using the python vpi example code provided but only at lower resolutions. At higher resolutions the results are very noisy. I am using cuda as the backend.

The rectified input image size is (4112, 3008) [width, height], After scaling it down to (1280, 720) or (1920, 1080) I get good enough results. But as soon as I use higher resolutions like (2560, 1440), (3840, 2160) or the input image size the disparity map generated is very noisy. Could you please advise on how could I go about fixing this?

[AastaLLL]

Hi,

Have you tried our latest VPI 2.3.9 in JetPack 5.1.2?
There is a new parameter called uniqueness can help the noisy problem.
https://docs.nvidia.com/vpi/group__VPI__StereoDisparityEstimator.html#structVPIStereoDisparityEstimatorParams

Thanks.

[not_a_bot]

Hi @AastaLLL , I am using the latest version of VPI and JetPack. The parameter does not help since at low resolution the disparities that are visible quite well become very noisy at higher resolution. I have tried various values of the parameters but not have gotten expected results.

I am also unable to upload any attachments for some-reason, hence I have messaged you the drive folder link. Please have a look and let me know if you need anything else. Thanks

[AastaLLL]

Hi,

Somehow I didn’t get your private message.
Could you share the link again?

In our previous testing, setting the uniqueness=0.7 or 0.8 can help with the noisy issue.
Thanks.

[not_a_bot]

Hi, I messaged you again but I am also sharing the drive_link here. If possible, it would be great if you could share your email-id as I could share my code and more information over there. Setting the uniqueness value 0.7 or 0.8 I get a almost black image meaning the disparity map generated is all noisy. Please have a look at the disparity maps shared at different resolution. I think it this has to do with the window_size param which is fixed and too small for higher resolution images.

[AastaLLL]

Hi,

Thanks for the results.

We are checking this with our internal team.
Will let you know the following.

[not_a_bot]

Hi @AastaLLL, is there any updates you could share?

[not_a_bot]

Hi @AastaLLL , just wanted to check in and ask if there are any updates you could share? Also please let me know if you need any more additional information

[not_a_bot]

Hi @AastaLLL , it would be great if you could share some updates? If you need any more information from my end please let me know

[AastaLLL]

Hi,

Sorry for the late update.
Our internal team is still checking on this.

Thanks.

[AastaLLL]

Hi,

Here is the update from our internal team:

The parameters that might improve (reduce) noise are: P1, P2, and uniqueness.

If you set the same uniqueness on low and high resolutions, please change P1 and P2 as they relate to the pixel neighborhood.
Moreover, it also recommended to try uniqueness from -1 (disregarding) to 0.90~0.99.

Thanks