Does non-planar region mask quality deteriorate after finetuning for depth estimation?

[MichaelRamamonjisoa]

Hi,

Thanks a lot for your great work! I was curious about your experiments on depth estimation.
One thing I noticed on PlaneNet is that when they finetune on NYUv2 with the depth loss only, they get less and less regions detected as planar (only the layout).

I was wondering if you observe a similar phenomenon in your case, and if it explains the depth performance improvement.
Could you maybe share some snapshots of [non-planar mask, instance mask, depth maps], before and after fine-tuning with your loss in Equation 10.

Thanks in advance for your help!