How to merge Semantic segmentation and Part segmentation?

[moddent]

Hi, there:

I want to merge Semantic segmentation and Part segmentation to generate the new label in .png format.

Can you give me some suggestion or some references.

Thx.

[pmeletis]

Hello moddent,
Currently, there is documentation to merge semantic and instance segmentation, and panoptic and part segmentation, but not semantic and part segmentation. We are planning to provide a script for that as well, so the semantic-part ids (sids_pids) can be generated and evaluated. For now, you can use the docs here by encoding your semantic segmentation as panoptic segmentation. @DdeGeus other suggestions?

[DdeGeus]

Hi @moddent, thanks for your interest in our work!

For what purpose do you want to use this exactly? If you want to generate just the semantic segmentation (stuff) and part segmentation (part) labels from our ground-truth, you can just use the sids and pids only when decoding the ground-truth as described here.

If you wish to merge semantic segmentation and part segmentation predictions as a 3-channel PNG as described here, you can simply treat all scene classes as stuff classes, as @pmeletis suggested, and follow the entire merging procedure. For merging to the 3-channel format, you just have to change the stuff class definition in the EvalSpec described here.

Does this answer your question?

[moddent]

Okay, I see.
Thanks for your reply.
Since I saw this and I'm just curious!