DekuLiuTesla / CityGaussian

CityGaussian Series for High-quality Large-Scale Scene Reconstruction with Gaussians
https://dekuliutesla.github.io/CityGaussianV2
Other
507 stars 36 forks source link

Too many folaters #55

Open DuHao55 opened 6 days ago

DuHao55 commented 6 days ago

Thank you for your outstanding work, I have successfully run through my own bridge pier dataset (close to 5000 sheets) using the parameters configured by rubble with the following results: image However, in the generated model, I noticed that the surface of the bridge pier appeared to have a lot of floaters, which resulted in a non-smooth surface and also resulted in the reconstruction. This is the image I captured with the robot, captured closer to the surface of the bridge pier(On the robot I used light boards, resulting in equal lighting for all images). _h0_l1 This is what I got from my rendering. 00000 The results don't seem to work very well for weak textures and there are too many floaters. What is the solution to this, please? Can I make it work better by adjusting the parameters?

yzslab commented 5 days ago

I think you can try smaller the LR of the xyz, and disable the opacity reseting.

DuHao55 commented 5 days ago

Sorry, I am a novice. Do you mean to set a smaller position inti LR? Does Disable the opacity mean I just need to delete these two lines of code in train_large.py?https://github.com/DekuLiuTesla/CityGaussian/blob/11ece7d88627e3a85f0979b502dd500c526f4367/train_large.py#L164

yzslab commented 5 days ago

Sorry, I am a novice. Do you mean to set a smaller position inti LR? Does Disable the opacity mean I just need to delete these two lines of code in train_large.py?

https://github.com/DekuLiuTesla/CityGaussian/blob/11ece7d88627e3a85f0979b502dd500c526f4367/train_large.py#L164

DuHao55 commented 5 days ago

I have retrained the model as per your comments and turning down the init and final LR did improve the model performance (PSNR improved by 1.5). And when I continued to make disable the opacity reseting, the model performed a little and I observed that it seems to floaters more and biger. 00000