search

hbb1 / 2d-gaussian-splatting

[SIGGRAPH'24] 2D Gaussian Splatting for Geometrically Accurate Radiance Fields
https://surfsplatting.github.io
Other
1.57k stars 77 forks source link

Many surfels are like spikes #56

Open CanCanZeng opened 1 month ago

CanCanZeng commented 1 month ago

I used the code in issue https://github.com/hbb1/2d-gaussian-splatting/issues/55 to check the model generated by 2D-gaussian-splatting, and found that it has a common feature with the result of 3DGS, that is, many primitives are like spikes, rather than more intuitive ellipses. I don't know if the authors has discovered this phenomenon. I wonder if we can add a length-to-width ratio constraint to the scales of gaussian, which may lead to better reconstruction of the scene.

hbb1 commented 1 month ago

I think it will. I construct an extreme case.

without constraint image

with constraint (su==sv) (isotropic surfels) image

You can see that the distribution of isotropic surfels looks better. and there are less artifacts near the arm of the chair. Although such constraint may harm the PSNR, it has the potential to improve geometry. Or some soft constraints as you mentioned can be used for further trade-off. This is from my thought, not experimentally validated. Feel free to discuss.

CanCanZeng commented 1 month ago

Your experimental results are very encouraging, and I think we can delve deeper into this area