Ignoring multiple classes when using SegNet

Autonomous Machines Jetson & Embedded Systems Jetson Nano
1

Hello,

I am creating an autonomous racing model that will go on a go-cart sized vehicle and ran laps around a track.

I am using Segnet on the Jetson nano to detect the roads. This model also detects a lot of other things.

My question is, can I ignore multiple classes using Segnet on only one using:

net.Process(img, width, height, “ground,void”)

Thank you for any help.
Orestis

2

Hi Orestis, currently there is only one class that it’s able to ignore. Note that this class will be ignored during argmax classification, so pixels that would have been classified as this class will instead be classified as the next-closest class. Hence using this with a popular class may lead to undesired mis-classifications. So perhaps instead what you want to do, is change your class labels and colors of the classes you don’t care about to all be the same.

Then in your processing routine of what uses the segmentation map for navigation, your routine can more easily ignore a certain set of classes/colors and just find “the road” or whatever it is you’re looking for. Alternatively, re-training the model with your desired output classes may lead to better performance.

3

Thanks Dusty,

That sounds like it could work.

Sorry if this is a bit obvious but,
where can I change the class labels and colors then?

Thanks,
Orestis

4

You would change them in the class labels file and colors file of your model. There should be these text files that go along with the model.

5

Thanks Dusty,

Its working greately now.