Skip to content

Question about input channels and semantic conditioning in MinkUNetDiff #1

@butterfly554

Description

@butterfly554

Hi, thank you for open-sourcing this excellent work! The diffusion-based semantic scene completion for point clouds is truly an interesting and promising direction. I have a question regarding the model architecture. While exploring the code, I noticed that in the instantiation of MinkUNetDiff, the in_channels is set to 3:

self.model = minknet.MinkUNetDiff(in_channels=3, out_channels=self.hparams['model']['out_dim'])

Given that the model performs semantic prediction, I was wondering: does the semantic information serve as an additional attribute/condition during the diffusion process, or is it handled differently in the network architecture? I would appreciate any clarification on how the semantic prediction is integrated into the diffusion framework. Looking forward to your insights!

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions