Some comments

[cornerfarmer]

• [x] The writer function should go into blenderproc/python/writer/SceneGraphWriter.py and should be available via bproc.writer.write_scene_graph. Its first argument should not be a file handle but just the filename, similar to the other writer functions

https://github.com/jplazag/BlenderProc/blob/e6e901ddac64a02f04ff584c16e22f6490658092/examples/automate_semantic_relations/third_placement.py#L357

• [x] The sampler function should go into blenderproc/python/object/SceneGraphSampler.py and should be available via bproc.object.sample_scene_graph.

https://github.com/jplazag/BlenderProc/blob/e6e901ddac64a02f04ff584c16e22f6490658092/examples/automate_semantic_relations/third_placement.py#L182

• [x] The coco annotations are only necessary to compute the bboxes right? I think it would be better that the write_annotations function compute the bboxes by itself using the given segmentation images

https://github.com/jplazag/BlenderProc/blob/e6e901ddac64a02f04ff584c16e22f6490658092/examples/automate_semantic_relations/third_placement.py#L702

• [x] We should add a verbose flag to the samples_poses_on_surface function instead of changing stdout:

https://github.com/jplazag/BlenderProc/blob/e6e901ddac64a02f04ff584c16e22f6490658092/examples/automate_semantic_relations/third_placement.py#L74

• [x] Why is it necessary to disable+enable rigidbodies if they already exist? Shouldn't it be enough to keep the rigidbody of the surface enabled (it will stay passive) and always disable the rigidbody of the dropped object after the simulation

https://github.com/jplazag/BlenderProc/blob/e6e901ddac64a02f04ff584c16e22f6490658092/examples/automate_semantic_relations/third_placement.py#L98

• [x] This function can be removed right?

https://github.com/jplazag/BlenderProc/blob/e6e901ddac64a02f04ff584c16e22f6490658092/examples/automate_semantic_relations/third_placement.py#L399

[cornerfarmer]

Looks already a lot better!

• [x] This if is now obsolete

https://github.com/jplazag/BlenderProc/blob/main/examples/automate_semantic_relations/third_placement.py#L147

• [ ] To determine which object is a table, could we maybe use the category_id instead of the name? mapping should define the mapping between category id and category name

https://github.com/jplazag/BlenderProc/blob/main/examples/automate_semantic_relations/third_placement.py#L86

• [x] Your new example folder automate_semantic_relations should go into the advanced folder. Also please remove all unused scripts and rename the main one to main.py

• [x] Please use one parameter for the output path. Also the number of cameras can be determined based on the given data, so this should be no extra parameter. The debug parameter test_bboxes should also not be part of the final code https://github.com/jplazag/BlenderProc/blob/main/blenderproc/python/writer/SceneGraphWriter.py#L12

• [x] In this code there are a lot of lines that are coco specific and are not really necessary here. You can remove the first few ifs. Also I think a lot of the latter code is not necessary. You only need bounding boxed for each category_id right? Then just use the last for loop and loop over all category ids and then use the segmap to calculate the bounding boxes https://github.com/jplazag/BlenderProc/blob/main/blenderproc/python/writer/SceneGraphWriter.py#L89