hturki / suds

Scalable Urban Dynamic Scenes
MIT License
199 stars 16 forks source link

question about dynamic RGB #17

Closed tb2-sy closed 1 year ago

tb2-sy commented 1 year ago

Thank you for sharing the code. After running your code eval.py, I found that in the result of dynamic rgb of vkitti2 scene, there will be some things other than dynamic vehicles (mostly in static tree areas), what is the reason?

hturki commented 1 year ago

Shoot, I knew I forgot something as part of the release.

This is partially due to how transmittance is calcuated - even if "dynamic" sigma values along a ray are low, they will have a high weight when rendering the pixel color RGB if there is no other high-sigma values ahead of them. In practice when rendering visualizations for the paper I'd zero out all of the weights along a ray when rendering the dynamic component when the accumulated DYNAMIC_WEIGHT value fell below a certain threshold.

Other hyperparameters that affect the static-dynamic factorization are the entropy loss term, the dynamic loss term, and the entropy skewness. But I'd probably start with just zeroing out pixels below a certain dynamic threshold when rendering first.