Resulting labels of function nppiLabelMarkersUF_8u32u_C1R_Ctx seperate connected component?

Accelerated Computing GPU-Accelerated Libraries
1

Hello,

I wanted to use Nvidias NPP library to find connected components, but from my perspective the results of the function nppiLabelMarkersUF_8u32u_C1R_Ctx are sometimes wrong.

The generated label markers seperate a connected region which means that neighbouring pixels of the same component have be assigned with different label markers (see example underneath).

The behaviour can be reproduced by running the example code …/7_CUDALibraries/batchedLabelMarkersAndLabelCompressionNPP/… several times and comparing the saved output file “Lena_LabelMarkersUF_8Way_512x512_32u.raw”. I only looked at the ROI: x=83, y=288, width=106, height=293. I guess, in 1 out of 10 runs the problem occurs.

Example: top (original), bottom(LabelMarkers): Imgur: The magic of the Internet
The left image is correctly filled with the same value but the right isn’t (yellow rectangle).

Has anyone observed a similar problem?

Thanks!

1 Like
2

I’m having a same problem, do you know how to get around this?

3

I have no solution for you, but I filled a bug report for the problem. According to the release notes it was fixed in CUDA 11.5 (" * Resolved an issue with nppiLabelMarkersUF_8u32u_C1R column processing incorrect results.").

Which CUDA-Version do you use?

4

Thanks.
I use CUDA 11.1