Large Scale Mean Shift Takes a Long Time

I ran the LargeScaleMeanShift tool using the application launcher (7.2.0-Win64) on a small test image (11MB, 2158 X 1428 pixels). This was on a 16 core system and it took 30 minutes. The log indicated that up to 16 cores could be used.

Is there a setting I’m missing to make sure this runs on multiple cores?

I also saw another post about a 1.45GB image crashing the algorithm from Oct 2020. I’m looking to run this on a 6GB Worldview image, so I hope it can handle it.

Hi @acanion ,

The LargeScaleMeanShift uses the streaming mechanism: that means that you can process images of any sizes.

Regarding processing time, it depends on the parameters of the segmentation. When large radius must be analysed around pixels, it usually takes more time.

Also, maybe changing tilesize and/or ram could help (I am not quite sure in this point).

Hope that helps!

Thanks for the response @remi.cresson. I tried the mean shift algorithm in ArcPro and it only took 2.5 minutes on the same image (using 2 cores). I’m not sure if these algorithms are comparable, even though they have the same name. I’ve found ESRI documentation to be pretty opaque.

In general, I get a better looking segmentation with Orfeo Large Scale Mean Shift than ArcPro, but if I can’t get the runtime down, I won’t be able to use Orfeo.