custom head for PTV3 semantic segmentation

[naisarg-pandya]

I have clone the point transformer V3 git repository (git clone https://github.com/Pointcept/PointTransformerV3.git). because i need to integrate this model with another model. here available model output feature shape (output["feat"] ) is (Nx64) so, do i need to designed any spacial head for sementic segmentation task or i can just put one linear layer and make the shape (N x num of class).

if i need to design spacial head for the task, can you suggest me some output head architecture which can probably give me best result on PTV3 model.

Please guide me for the custom training on the custom data for PTV3 model.

[Gofinge]

Hey, just for semantic segmentation, I think one linear layer is alreay enough?