search

google / dynibar

Implementation of DynIBaR Neural Dynamic Image-Based Rendering (CVPR 2023)
https://dynibar.github.io/
Apache License 2.0
851 stars 40 forks source link

Question about how to get a disparity maps #7

Open otonari726 opened 1 year ago

otonari726 commented 1 year ago

First of all, thank you for your great work and for publishing the code!

I’m trying to experiment with my own data and was wondering how to get the disps.

My understanding is that you need to create checkpoints in train.py before running test.py. Did you run train.py using own data and then test.py?

Also, can the disparity maps be replaced by the results from MiDas used in Neural Scene Flow Fields?

Thank you again!

zhengqili commented 1 year ago

Hi,

Yes. You need to perform per-scene optimization by running train.py before running test.py for dynamic CVD models. We do tried use idea from NSFF with scale-shift in variance loss for depth supervision, but found training to be much less stable. Therefore, we instead use depths estimated from CVD methods which provide scale consistent depths so that we can directly apply L2 or L1 losses on the depths.