Replicator composer creates bounding box for all objects

Hello. I am trying to generate synthetic data using replicator composer. I am only interested in creating bounding box around single object which I am interested in.

However, when I run the replicator composer, it gives me bounding boxes around every unique object present in the scene (Image attached in this post). I only want to generate bbox around the brown box that’s all. Is that possible ? If yes, then please guide me what I am doing wrong. If no, then is there any other walk-around ?

I am attaching the yaml file I am using to generate synthetic data below.

sample_2.yaml (1.1 KB)

Isaac Sim Version: 2022.1.0

In the bounding box NPY output files, you can filter each box by its name. The boxes corresponding to the scenario, in this case the scenario room (floor, walls, and ceiling), the group name is “[[scenario]]”.

I am trying to train an object detection model which requires labels to be in KITTI format. Based on what I have learned for Replicator Composer, I can only get output for 2D bounding box as NPY and RGB with bounding box under “visuals”. Is there any method through which I can get the “.txt” output for bounding boxes ?

hi @mayank.ukani

Did you find out the answer to this ?
I guess it would not be too hard to extract the label data from the .npy files- but I’d prefer to avoid this extra step if possible

thank you


Yes, we can extract labels and their respective bbox co-ordinates and create KITTI expected .txt labels by creating a simple python script. If Replicator Composer could directly give us labels in KITTI format then it would be less hassle. I guess that is the only way to go.


yes, that is what I’ve done- although the npy file does not seem to contain all of the data needed for the KITTI format
I can see class name & bbox x/y min/max values- but the occlusion and truncation fields would be really useful to me

I’ll have a look through the Replicator Composer python tomorrow, but I’d have thought that RepComp would be able to produce labels in this format


Nvidia’s Pre-trained Object detection models only support class names and bbox co-ordinates as of now. So you can pad other parameters with zero. Even if we would have gotten values for other parameters, we cannot use it as of now.

Also, please share your python script (if you have written) to extract labels and their bbox co-ordinates. Thanks!


This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.