Topics created by paul.foley

Hi @Riccardo-Benedetti and @paul-foley ,

This issue is not common. There is a potential condition (which we are going to look into) when a MPA client crashes and does not close the pipes properly, the MPA server side will keep the allocated resources open. If the client crashes repeatedly, eventually there will be no more memory for the Kernel to allocate for the new pipes.

Please check to make sure that you have no client processes that are subscribing to image streams and are crashing and restarting (perhaps restarting automatically by systemd). It is possible that a process that is not your test app is misbehaving for some reason and is causing this memory leak by continuous crashing and re-starting.

Also, what if you just use voxl-inspect-cam to inspect the single camera stream and no other camera clients running, does the same issue happen?

Alex

@tom that worked, thank you!

@paul-foley , we do not have a repository for images.

My suggestions would be:

try using training sets from other sources (as you have been trying), but I don't have a specific reference in order to better match the 3rd party datasets to your use case, you should try to do the following: transform the 3rd party images to make them look more similar to what you are looking for (crop, zoom, resize, adjust color, etc). This can be automated transform the images collected on Starling to some standard, for example you can perform fisheye un-distortion to make them look more like the images taken from non-fisheye cameras from your training set, similarly adjust color if needed (white balance).

In general, if you can somehow standardize your images, and convert the 3rd party images to the same standard, you can greatly benefit from the large data sets that you could find online. The only down side that when running the model online, you will need to transform your images in real time, but that should not be a big deal. If you need to, I can help you with some GPU-based examples to do image transformation very quickly, but you should first make sure that this approach will work for you first (and optimize later 🙂 )

I hope this helps..

Alex