Sonata pre-training details

[delyan-boychev]

First off, congratulations on the great work! The paper is very well-described.

I have a question regarding the usage of Sonata in your experiments. I noticed that the accepted features differ from the original Sonata implementation, which suggests it was likely pre-trained (on semantic segmentation task before the pre-training of the SpatialLM) on your specific training set.

Could you clarify if this is the case? If so, would it be possible to share the pre-trained checkpoint or the configuration file used for pre-training?

Thanks in advance for your help!

[SamMaoYS]

Hi @delyan-boychev, thank you for your interest.
No we didn't finetune the Sonata, we use the pretrained Sonata backbone, and drop the input weights for normal channels.