Reasoning behind not keeping widht and length scale in density maps?

[Grantus1407]

Hey, I am trying to use this model as a basis for my master thesis and i noticed after taking the Structured3D Dataset to try and get the semantic information out of the scenes, that the density maps provided by you are always scaled to perfectly fit the 256x256 size of your images and thus lose the real scale of the floorplan.

I wanted to know if this was a deliberate decision to use most of the image space as a feature or if this was an oversight.
And how the usage would be for a real life scenario? Would you save the compression value somewhere and "uncompress" the floorplan after creation?

[ywyue]

Hi @Grantus1407, this design choice is consistent with our baselines, MonteFloor and HEAT. I think the motivation is to use most of the image space. However, I didn't ablate this, i,e, comparing original input V.S. scaled input for training. We save the scale values and unscale the prediction back to the original scale, as indicated in all of the visualizations in our paper.