Question about the feature dimension when rendering 3D Gaussians to the BEV plane

[captainyhj]

Hi:
thank you for your excellent work!
I found the code for rendering 3D Gaussians to BEV features, as shown below:

GaussianLSS/GaussianLSS/model/GaussianLSS.py

Lines 175 to 184 in cd6d305

	rendered_bev, _ = self.rasterizer(
	means3D=means3D[i][mask[i]],
	means2D=None,
	shs=None, # No SHs used
	colors_precomp=features[i][mask[i]],
	opacities=opacities[i][mask[i]],
	scales=None,
	rotations=None,
	cov3D_precomp=cov3D[i][mask[i]]
	)

The dimension of the rendered BEV feature is [128, 200, 200], where the embedding dimension is 128. At first, I thought this dimension was determined by the dimension of the input feature(in the line 179: features[i][mask[i]]) . However, when I changed the dimension of the input feature to 64, the dimension of the rendered BEV feature was still [128, 200, 200], which confused me a lot. What determines the embedding dimension of the rendered BEV feature?

[shu-wei-lu]

@captainyhj , since the gaussian splatting implementation does not support changing features dimension dynamically, if you pass in a feature that is not the same as dimension when compiled, which is 128 here, there would be a bug. If you would like to change to different embedding dimension, you can edit this parameter here:

GaussianLSS/GaussianLSS/model/diff-gaussian-rasterization/cuda_rasterizer/config.h

Line 15 in cd6d305

#define NUM_CHANNELS 128