google/dynibar

Question about the data preprocess

CTouch opened this issue · 2 comments

CTouch commented

Hello, thank you so much for making your work open source.
I am trying to use other data from DAVIS. But the source virtual view created by render_source_vv.py is bad.
This is the original image.
00018

This is the virtual view image.
06

But I think the model reconstructed by colmap is good enough.
train_colmap_reconstruct

I wonder if there is some mismatch between colmap output and dynamic-video-depth. Or could you please give me a pair of input and output for colmap and dynamic-video-depth to help me find out my mistake.

Thank you a lot!

zhengqili commented

Hi, the virtual view image looks OK to me. Why do you think it's bad?

CTouch commented

Thanks for your reply!
02
This is virtual view image provided by you. I think it may be closer to the real situation, which reflecting the depth and disparity.
And the image resolution seems to be lower.

https://github.com/google/dynibar/assets/65765708/b45da9a1-015a-4fee-a598-146434851a7b
I try to set the novel view camera offset fixed to be (0, 0, 0.1). But the output is not as good as your demo. I am trying to find out why.

Maybe I should check the NSFF part?