[question] Any suggestions on normalizing the outputs better?

[wes-kay]

Original:

Output:

This is currently an image that I took of a port hole in converted to ply, using the .pfm, and also normalized it before converting, viewed in blender.

The issue I'm having is realistically the edges should be flatter and not scale to infinity like it currently is, is there any real tricks to making the pfm a little better to work with?

Any suggestions are appreciated.

[northagain]

disp to depth with min and max depth
using this code may be helpful
def disp_to_depth(disp, min_depth, max_depth):
    """Convert network's sigmoid output into depth prediction
    The formula for this conversion is given in the 'additional considerations'
    section of the paper.
    """
    min_disp = 1 / max_depth
    max_disp = 1 / min_depth
    scaled_disp = min_disp + (max_disp - min_disp) * disp
    depth = 1 / scaled_disp
    return scaled_disp, depth

[isJHan]

disp to depth with min and max depth using this code may be helpful def disp_to_depth(disp, min_depth, max_depth):     """Convert network's sigmoid output into depth prediction     The formula for this conversion is given in the 'additional considerations'     section of the paper.     """     min_disp = 1 / max_depth     max_disp = 1 / min_depth     scaled_disp = min_disp + (max_disp - min_disp) * disp     depth = 1 / scaled_disp     return scaled_disp, depth

There is no 'additional considerations' in the paper I downloaded at IEEE website. Can you provide the right version? Thank you!

Besides, I'm confused about 'sigmoid output' in your code. In the dpt_beit_large_512 network, I can't find the sigmoid layer. I find the output of my network is between 900 and 10000. In this case, how can I transform output to depth?

Thank you very much!

[thucz]

I have the same question. The inverse depth output is between 900 and 10000. Is the problem resolved?

[isJHan]

I have the same question. The inverse depth output is between 900 and 10000. Is the problem resolved?

Hi. I convert the output to depth map by this way. Firstly I inverse the output directly by ‘depth=1/output’, then use min-max normalize method ‘depth= (depth-depth.min())/(depth.max()-depth.min())’ for an valid depth map.

[thucz]

@isJHan Thanks!

[isJHan]

@isJHan Thanks!

Today I found a bug in this procedure. When I infer on another dataset, the output can be negative or 0. So there will be a bias added to output like depth=1/(output+bias). Do you have a better way? Thanks!

[thucz]

Now I also add a bias and a scale to the output like this (but not to [0, 1]):

https://github.com/KU-CVLAB/DaRF/blob/47b2d1a23d13f0d149e55cf8fd2195ec42093d1e/plenoxels/models/dpt_depth.py#L87C18-L87C18

[isJHan]

Now I also add a bias and a scale to the output like this (but not to [0, 1]):

https://github.com/KU-CVLAB/DaRF/blob/47b2d1a23d13f0d149e55cf8fd2195ec42093d1e/plenoxels/models/dpt_depth.py#L87C18-L87C18

Thanks. But how can we get alpha and beta for another dataset?

[thucz]

@isJHan
See the supplementary of RichDreamer in Page 13-14 (Sec A.2). It shows the general normalization methods