Prune points
bmikaili opened this issue · comments
Has anyone tried combining this approach with a prune step like in https://github.com/VITA-Group/LightGaussian?
This is an interesting idea. I try to set a more strict pruning rule during training to simulate the LightGaussian (not perfectly consistent). For simplicity, I set self.min_opacity
to 0.005
(original), 0.2
, 0.5
. The results on Tandt/truck
are as follows:
![image](https://private-user-images.githubusercontent.com/9450147/289791272-65504e14-54d7-4cbe-8ae7-ee91615cfdb0.png?jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.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.50nYzjasu5s9eIwmYN3Az9i9vfa8zhLbuCNwmpsJGhk)
With a more strict rule (larger min_opacity), Scaffold-GS enjoys a smaller storage and faster speed, at the sacrifice of LPIPS. This is just a simple and not very rigorous exploration, further conclusions are welcome in the future.