jeffwang987 / movedepth Goto Github PK

I have a stupid question when I meet the first sentence:
“Self-supervised monocular methods can efficiently learn depth information of weakly textured surfaces or reflective objects.”
Which method solved those challenges?

Hi, thanks for your nice work!!!
I'm a new guy for this mono depth domain. I just have one primary question. how can we align the depth with real world scale for follow-up task?

Thanks for the contribution of the code and paper. It is inspiring and helps the community forward.

From the initial version, it seems that two minor bugs are presented in code:

Can not import MovedepthOptions:

It seems to be a minor bug when cleaning the code:

• probably update here with

   from movedepth.options import MonodepthOptions
  

• probably update here with

   options = MonodepthOptions()
  

Gradient error in newer PyTorch

This is raised because newer pytorch is more restricted to inplace modification that changes the gradients.

Probably change here into

  out = out + x

Then as tested, the code can run at torch==1.11.0

Hi,

Thank you for sharing your project! It is really amazing.

It seems that the evaluate_depth.py provides the evaluation for mono depth and mvs depth, and upper bound results, but lacks the evaluation for fused depth. I tried to add the evaluation code for fused depth only to find that fused depth is worse than mvs depth with my evaluation code. The test results with your pretrained weights are as follows:

mono results:
abs_rel | sq_rel | rmse | rmse_log | a1 | a2 | a3 |
& 0.113 & 0.823 & 4.724 & 0.190 & 0.879 & 0.960 & 0.981 \

mvs results:
abs_rel | sq_rel | rmse | rmse_log | a1 | a2 | a3 |
& 0.094 & 0.704 & 4.389 & 0.175 & 0.902 & 0.965 & 0.983 \

fuse results:
abs_rel | sq_rel | rmse | rmse_log | a1 | a2 | a3 |
& 0.112 & 0.818 & 4.709 & 0.190 & 0.879 & 0.960 & 0.982 \

upbound results:
abs_rel | sq_rel | rmse | rmse_log | a1 | a2 | a3 |
& 0.082 & 0.618 & 4.135 & 0.162 & 0.915 & 0.969 & 0.985 \

I also tested my own trained model, the results are as follows:
mono results:
abs_rel | sq_rel | rmse | rmse_log | a1 | a2 | a3 |
& 0.116 & 0.920 & 4.868 & 0.193 & 0.874 & 0.959 & 0.981 \

mvs results:
abs_rel | sq_rel | rmse | rmse_log | a1 | a2 | a3 |
& 0.097 & 0.785 & 4.512 & 0.177 & 0.899 & 0.963 & 0.982 \

fuse results:
abs_rel | sq_rel | rmse | rmse_log | a1 | a2 | a3 |
& 0.098 & 0.788 & 4.518 & 0.177 & 0.899 & 0.963 & 0.982 \

upbound results:
abs_rel | sq_rel | rmse | rmse_log | a1 | a2 | a3 |
& 0.086 & 0.698 & 4.269 & 0.165 & 0.911 & 0.968 & 0.984 \

I think there must be something wrong with my modification of the evaluation code. Could you please update the evaluation code to test fused depth?

How do we figure out the camera frame rate? I couldn't find anything related in the code.