questions about hybrid representation

[zhanghaoyu816]

Hi, Thanks for your nice project. I have some questions about your paper.

1. I would like to ask how you observed the phenomenon that Gaussian primitives tend to increase in size rather than optimize their positions under sparse view settings. Are there any visualization experiments to support this conclusion?

2. I would like to know how ray-based structured Gaussian primitives address the issue mentioned above. My confusion regarding this problem is that structured Gaussians can optimize not only the z_val (position) but also the scale and opacity. So, how can we be certain that this design prevents Gaussians from simply increasing in size? It seems that the design merely restricts the direction of position changes, but does not enhance the tendency of Gaussian primitives to correct their positions.

3. What are the differences between “Triang. init + 3DGS” and “w/ Hybrid Rep”? Why does the latter achieve even lower results than the former? It seems that the primary performance gains actually stem from the denser initialization obtained through the match priors rather than the optimization design for z_val.
   

I‘m looking forward to your reply, Thank you.

[prstrive]

Thank you for your interest in our work!

1. We analyze this from the perspective of visualization of the rendered image and not yet from Gaussian primitives. And judge from the overly smoothed results that the Gaussian primitives are increasing in size.
2. Because the shape and position of the Gaussian primitives are jointly optimized in 3DGS, and we use the ray-based Gaussian primitive to explicitly optimize the position, which can mitigate the risk of sub-optimization to some extent.
3. "Triang. init + 3DGS" means directly using the triangulation points to initialize the 3DGS. and "w/ Hybrid Rep" means only using the hybrid representation (including non-structure and ray-based). And the reason why "w/ Hybrid Rep" is worse is "w/ Hybrid Rep" haven't optimize the position of ray-based Gaussian primitives yet.