The color value of rendered image will exceed 1.0

[wzy-99]

Hello, I found that the color of predicted gaussian splatting will exceed 1.0.

I test re10k data, and add a break point here:

NoPoSplat/src/model/model_wrapper.py

Line 231 in 5af8e16

output = self.test_step_align(batch, gaussians)

I run the following code to check the maximum color value:

output.color.max()
tensor(1.0807, device='cuda:0')

The max color is 1.0807 > 1.0

What caused it? Is it because the coefficient of SHs is not being constrained?

And how to avoid this?

[botaoye]

Hi, thanks for pointing this out. I am directly using the RGB values rasterized by this cuda implamention. I am not sure if this is the expected behavior, I will check further and let you know.