search

autonomousvision / monosdf

[NeurIPS'22] MonoSDF: Exploring Monocular Geometric Cues for Neural Implicit Surface Reconstruction
MIT License
573 stars 53 forks source link

The estimated depth results of ScanNet via omnidata model #4

Closed lzhnb closed 2 years ago

lzhnb commented 2 years ago

Thanks for your amazing work!

As you said in the paper, you leverage the omnidata model to estimated both monocular depth and normal. We've downloaded the pretrained provided by the official, and it really performs well in the details. However, we found that the range of the estimated depth is so small. Here is the result of one image from 0616 scene in ScanNet. c12c2e276c3e072c1198b4d33fce781

In our acknowledge, the range of depth in ScanNet should be 0~3. How do you process such estimated depth?

Good luck :)

niujinshuchong commented 2 years ago

Hi, we use scale-invariant loss to use omnidata depth so the range of output dosen't matter.

jzhzhang commented 2 years ago

@niujinshuchong
I also notice that your code recommends using ominidata to obtain the depth and normal for custom datasets. I wonder if there will be a performance drop when we instead use real depth maps from depth sensors.

niujinshuchong commented 2 years ago

@jzhzhang If you have real depth map then you don't need to use scale-invariant loss. You could use MSE or L1 loss directly.