Genearting ground truth object masks

Computer Vision / Perception
1

Hi,

I am looking at creating a ground truth mask for a camera simulated in gazebo or from the URDF is there an existing method for this?

Like this:
image

thanks,
Guy

2

Heads up, I realize this is a question that could reside on ROS answers, but I also know that the answer, is a bit more long for than might be appropriate for ROS answers, and it is a good discussion question. Anyway, I think it is a really good discussion question. I know there was a GSoC project in this direction a few years back but I don’t have much more information than that.

3

Thanks. I did think about it being a ROS answers question, but I assumed there was no assumed correct answer yet, and there are multiple interesting ways that this could be achieved.

4

Like I said I’ve heard that there was some Gazebo work in this direction by a GSoC student a few years back but I have yet to see it. You may want to comb through past GSoC posts.

1 Like
5

I am not aware of gazebo providing semantic/instance segmentation camera. Depending on what kind of envoirment you want to simulate I can suggest 2 places to look up ;

Nvidia’s NDDS: https://github.com/NVIDIA/Dataset_Synthesizer
CARLA: https://github.com/carla-simulator/carla

NVIDIAs NDDS is very much suitable to any object, as long as you have cad models, then you can quickly generate thousands of images with GT labels.

CARLA is more suited for Autonomous driving scenarios.

both of these are built on top of UnrealEngine

6

Is there a way to get SRDF or URDF into CARLA or NDDS?

7

don’t think so , they accept cad model file formats. e.g .obj.

8

Yeah, looks that way, and there seems to be no GSOC’s that cover this either.