Rendered RGB images and semantic maps

[DingYikang]

Hi, I noticed that the Tab.6 of the paper (arxiv version) mentioned that the photometric loss doesn't improve the performance. However, I am wondering how the rendered RGB and semantic maps look like. Would you provide some visualization results?

Thanks!

[huang-yh]

Sorry for the late reply. The rendered RGB images are quite blurred which might be related to the supervision (l2 loss only) and inappropriate hyperparameters (number of Gaussians, etc).

[junho2000]

@huang-yh
Hi Huang,

Thank you for providing additional result about photometric loss
Btw, i am just curious about how you render RGB.
Did you just project gaussian points like vanilla gaussian splatting? like containing every gaussian points have spherical coeff or just using RGB MLP from voxel representation to predict the color?