BOP class and instance segmentation maps for object textured surface sampling, pose sampling and physics positioning and challenge 2020 data generation

[preethamam]

I am able to generate the color and depth images, scene_camera.json, scene_gt.json and camera intrinsics using camera.json of BOP HOPE objects on textured surface sampling and object pose sampling and physics positioning programs.

I am wondering how can the class and instance segmentation maps can be generated for the same scenes using the main.py files of BOP object textured surface sampling, pose sampling, and physics positioning, and challenge 2020 data generation codes.

Lastly, after obtaining the depth maps of each color/RGB image, is there a way to generate the corresponding 3D point clouds with class and instance segmentation labels? In addition, how can I use --num_scenes=2000 option for the object textured surface sampling, and pose sampling and physics positioning class of problems? I get the below error:

usage: main.py [-h]
               [bop_parent_path] [bop_dataset_name] [cc_textures_path]
               [output_dir]
main.py: error: unrecognized arguments: --num_scenes=2000

Please note that I have installed bop_toolkit.

Thanks,

[MartinSmeyer]

Hi @preethamam

I am wondering how can the class and instance segmentation maps can be generated for the same scenes using the main.py files of BOP object textured surface sampling, pose sampling, and physics positioning, and challenge 2020 data generation codes.

See here

Is there a way to generate the corresponding 3D point clouds with class and instance segmentation labels

Yes, it's easy, just use the inverse of the intrinsic camera matrix. It's not part of BlenderProc, but you will find a lot of info online.

--num_scenes=2000

This is just available for the bop_challenge examples. You can easily adapt the example for other use cases, but you need to make yourself familiar with BlenderProc.

[preethamam]

@MartinSmeyer thanks for your suggestions.