Regarding the Quality of Training Results

[LuXiao0228]

When I train with sh_degree=0, the training results are much worse when using taming 3dgs' Rasterizer compared to using the original 3DGS Rasterizer. May I ask if this phenomenon is normal?

[rahul-goel]

Hi.

I tried this with the bicycle scene until 7K iterations. I see similar metrics. Can you give the steps to replicate your issue?

Rahul

[dingjr7]

Hi, i tried the 5aef77d and trained the garden for 500 iters. The result is blur and has low psnr compared with gaussian splatting.

[rahul-goel]

At 500 iterations, densification has not started. The results are expected to be blurry.

Also, at 500 iterations the sigmoid activation should be used for rendering. The default render script has abs activation as default since it is meant to be used after 30K iterations.