Hi,

I would like to use Gazebo to produce a dataset to train an NN-based autoencoder to perform semantic segmentation (ex: https://www.cityscapes-dataset.com/).
The ideal output format for doing so would be that the resulting image is a mask where the pixel value is the "number of the model" in the scene, or eventually separate channels for each model.

Let's say my world is a room with 4 walls a door and a plant inside So in the first case, the pixel value of the resulting image would range from 0 to 5 1- ground
2- walls
3- ceiling
4- door
5- plant
0- anything else

In the second case instead (multiple channels) the pixel value range will always be 0 or 1 but each of the 5 channel will only mask a specific model.

In case there is not such a plugin, does anyone have ideas or resources about to implement it?