YvanYin/VNL_Monocular_Depth_Prediction

Some questions about surface normal estimation and robutness test

mei123hao opened this issue · 3 comments

mei123hao commented

Hi YvanYin:
Thanks your excellent work. After reading your paper I have some questions?
(1)I found that you have robustness test for virtual normal and surface normal in the paper. how do you calculate the mean difference error ? can you release your code about your rubustness test? Thanks a lot !!
(2)In the code recover_surface_normal.py, I found you add Identity Matrix:
ATA = ATA + eps_identity
is this for avoiding singular problem when perform LU decompose when running
n_img, _ = torch.solve(AT1, ATA) ?
(3) Can i understand that the method you use to recover surface normal is based on the SVD decompose method?
My email is 809741329@qq.com, mayby we can communicate through email, :D !!!

YvanYin commented

Hi,

  1. which table or figure do you refer to?
  2. Yes, it amis to avoid the singular problem.
  3. This is a least-square fitting method.
mei123hao commented

Thanks for your replying ! The figure that i refer to is figure 5 in the paper, which is shown like following:
image

YvanYin commented

Hi,
The Mean Difference Error is: Err = 1/n sum(arccos (n_{i} * n_{i}^{*}) )