Triangulated depth with dynamic objects

[VitorGuizilini-TRI]

Hi, it's me again. I'm using your triangulation for depth estimation from optical flow, and these are some results I am getting on VKITTI2:

GT depth

GT optical flow

Triangulate depth

As you can see, it works fine everywhere except on dynamic objects (and some strange artifacts on the sky). Is that expected, or am I doing something wrong? I'm assuming you don't have that many dynamic objects on the datasets you evaluated on the paper, so that's why I am asking.

[MathLens]

Triangulation assumes the points are static across multi-view, so that is expected, which is also one of the reasons we have a depth refinement work based on single view after the triangulation.

[nauti16]

Hi, it's me again. I'm using your triangulation for depth estimation from optical flow, and these are some results I am getting on VKITTI2:

GT depth

GT optical flow

Triangulate depth

As you can see, it works fine everywhere except on dynamic objects (and some strange artifacts on the sky). Is that expected, or am I doing something wrong? I'm assuming you don't have that many dynamic objects on the datasets you evaluated on the paper, so that's why I am asking.

Hi, VitorGuizilini-TRI.
I followed your work using the VKITTI2 dataset, but the depth results I obtained are not as clean as yours. Could you share the method you used?

[VitorGuizilini-TRI]

Do you mean my triangulation results, like the one I'm showing above?

[nauti16]

Yes, regarding your triangulation result, below is the result I obtained by triangulating the VKITTI2 data. As you can see, the depth result isn't coming out very well.

Would any adjustments to the extrinsic parameters or other preprocessing steps be necessary?

[VitorGuizilini-TRI]

The code should be available on my repository here: https://github.com/TRI-ML/vidar/blob/main/vidar/utils/flow.py
Let me know if it works for you!

[nauti16]

Dear VitorGuizilini-TRI,
I have reviewed your code and confirmed that the extrinsic pose needs to be inverted.
I missed that in VKITTI2.0 dataset. Thank you very much!!!