Blenderproc: dataset creation

[soans1994]

Dear Authors,

i have some queries regarding synthetic dataset creation using blenderproc.
Which is the good format for dataset creation from 3d models. currently im generating coco style dataset as per https://github.com/DLR-RM/BlenderProc/tree/main/examples/advanced/coco_annotations
I have tried it for generating dataset from 3d sign board models. I want to build a larger dataset. However, when i have many 3d models of sign boards, what approach should i use. Should i build all 3d sign boards in a single blender scene and generate coco datasets(or do you recommend any other). Or is it possible to combine the coco annotations if i generate group of sign boards in a seperate scenes since i assign different obejct id for each sign board.

Thank you

[cornerfarmer]

Hey @soans1994,

I would recommend to perform multiple runs of BlenderProc and let them be merged into one coco_annotations output file. You can do this by enabling the parameter append_to_existing_output, like it is done here:

    {
      "module": "writer.CocoAnnotationsWriter",
      "config": {
        "append_to_existing_output": True
      }
    }

This will now keep the already existing output files and just append the newly rendered images. The categories in the coco_annotations.json file will be merged. Here it is important that the same categories have the same category_id and the same name across different runs, otherwise they will be recognized as different categories.

Let me know if you need any further help.

[soans1994]

@cornerfarmer
thank you very much. I will add "append_to_existing_output": True to config. Also, i should use the same output file name in the argument when running the coco annotations code. right?

[cornerfarmer]

Yes, the output_dir parameter has to point to the same directory.

[soans1994]

yes, understood. thank you

[themasterlink]

Great :) I am closing this.