3DGS-mcmc has big effect in scenes with large brightness changes?

[KotaYonezawa]

Thanks s lot for great code release!

I tried 3dgs-mcmc with my own dataset. (shot by Insta360 ONE RS 1-inch camera (https://www.insta360.com/jp/product/insta360-oners/1inch-360))
Then, I noticed that 3dgs-mcmc can remove terrible floaters that tends to appear in scenes with large brightness changes.
But I think such effect is not written in the paper.
Does anyone find similar phenomenon?
If possible, I would like to know why such big difference was occurred between vanilla 3dgs and 3dgs-mcmc...

vanilla 3dgs

bridge_3dgs_s.mp4

3dgs-mcmc

bridge_3dgs-mcmc_s.mp4

Note: This scene was shot from road only. (Drone was not used.) So the bridge does not look so good from the side.
(Blue line is camera path.)

[yzslab]

From you video, I think by reducing the learning rate of xyz, or increasing the densify_grad_threshold, you should be able to get a better result in vanilla 3DGS.

[KotaYonezawa]

@yzslab
Thanks a lot for your great advice!
I tried to change --position_lr_init = 0.000016, then floater was gone with vanilla 3DGS, as you suggested.
(Hmm... it's difficult to search best parameter value...)

vanilla 3dgs (--position_lr_init = 0.000016)