VPI1 Stereo Disparity - Confidence map is too high

[leo.chaumartin]

Hi,

I am having hard times with the VPI 1.X stereo disparity sample on my Jetson Nano B01. I made some virtual benchmark images :

Virtual cameras are perfectly aligned, with a subtle amount of grain to seem more realistic.
As a result with CUDA backend, I get this disparity map:

And this confidence map:

I have a few questions about this output:

• What is this black rectangle on the top-loft corner? It doesn’t show up with CPU backend.
• Why is the confidence so high in the sky?

I read this topic with a similar issue, but I cannot use the uniqueness parameter, as I am stuck with VPI1 / JetPack 4.6 on my Jetson Nano B01.

I am afraid that the confidence value just takes the correlation score into consideration.
What do you advise to get a usable confidence map ?

Thank you in advance.

Configuration:
Jetson Nano B01 4GB (eMMC 16GB)
Ubuntu 18.04
JetPack 4.6 (L4T 32.6.1)
CUDA 10.2.3
VPI 1.1.15

[AastaLLL]

Hi,

Have you tried tuning the parameter or downscaling the images to see if it helps?
https://docs.nvidia.com/vpi/1.2/group__VPI__StereoDisparityEstimator.html#structVPIStereoDisparityEstimatorParams

Thanks.

[leo.chaumartin]

Thank you for this quick reply.
I managed to change the confidenceThreshold parameter to its maximum (65280). It is slightly better, but doesn’t seem to solve the problem.

Downscaling is not available with CUDA backend : VPI_ERROR_INVALID_ARGUMENT: Only downscale factor of 1 is supported
I you meant manually reduce input size, it doesn’t change anything except the size of the black rectangle on the top-left:

By the way, this rectangle is probably due to the way I apply grain, since it doesn’t show up when I use real images. I’ll try to investigate this.

Thanks

[AastaLLL]

Hi,

Rescale can help you to downsample an image:
https://docs.nvidia.com/vpi/1.2/algo_rescale.html

The suggestion is from some of our users found the VPI stereo matching perform better in the low-resolution image.
So it’s recommended to give it a try.

Thanks.