search

autonomousvision / monosdf

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

About scene boundary #46

Closed lazyDamon closed 1 year ago

lazyDamon commented 2 years ago

Hi, The depth and normal information of the floor scene was lost，when I train high resolution real scene images. Could you please give me some advice? normal_710_18

niujinshuchong commented 2 years ago

@lazyDamon We also observe using multi-res. grids give noise results when we do experiments on scannet dataset. I think you could try to increase the weights for smoothness loss or try something like using large smoothness loss at the beginning of training and slowly decrease the weight for smoothness.

lazyDamon commented 1 year ago

@niujinshuchong We can see the scene that in the green box area is lost. It leads to difference between mesh and the real scene. I wonder if this is caused by the smooth? Or is it caused by the set boundary? In your project, you should normalize the scene into a cube. I want to know how you did it, and what is its scale? 2

niujinshuchong commented 1 year ago

Hi, the normalization is here https://github.com/autonomousvision/monosdf/blob/main/preprocess/scannet_to_monosdf.py#L75-L83

lazyDamon commented 1 year ago

OK, thanks for your great work!

Yzhbuaa commented 1 year ago

@lazyDamon Hello Damon, have you solved this problem? I observed the same thin structure loss on my own datasets. Could you share how to solve this problem? Thank you!