Green tint with splatfacto-big and bilagrad

[gradeeterna]

Splatfacto-big with bilagrad enabled has a green tint, not present with regular splatfacto + bilagrad, or splatfacto-big with bilagrad disabled:

splatfacto-big + bilagrad enabled:

splatfacto + bilagrad enabled:

splatfacto-big + bilagrad disabled:

splatfacto + bilagrad enabled + some arguments to match splatfacto-big looks fine also --pipeline.model.densify-grad-thresh 0.0006 --pipeline.model.cull-alpha-thresh=0.005 --pipeline.model.continue-cull-post-densification=False

@SharkWipf commented on Discord that the only other difference between regular and big splatfacto is the learning rate:

I've only tested with this tricky dataset where the exposure is all over the place, but can test some others soon.

Here is the ground truth image, and another over-exposed image from nearby:

[jb-ye]

@KevinXu02 Do you know why the learning rates of bilateral grid are different for splatfacto and splatfacto-big?

[abrahamezzeddine]

Out of curiousity, how do I enable bilagird with splatfacto? I would like to try this myself.☺️ thanks!

[jb-ye]

@abrahamezzeddine --pipeline.model.use_bilateral_grid True

[xiemeilong]

Could you please tell me the difference in training FPS when using bilateral_grid and not using it? I've noticed a significant slowdown in my code after adding bilateral_grid and I'm wondering if this is expected behavior.

[KevinXu02]

Sorry for the late reply. I forgot to update the lr rate for splatfacto-big when tuning splatfacto. So it should be 2e-3 for splatfacto-big.

The overhead for bilateral_grid should be scaled by the image resolution. There can be a 20% slowdown for 960p images, but may hurt more if using higer resolution images.