Question about Scale Normalization

[OrigamiStationery]

Thank you for your marvelous work! But I am confused: since utonia has designed Perceptual Granularity Rescale, is it necessary to perform scale normalization on the point cloud before encoding it with utonia, that is, scaling it onto a sphere of (-1,1)?

[Yujia-Zhang0913]

Thanks for your interest in our work.
The scale parameter is mainly related to the desired perceptual granularity. The purpose of Perceptual Granularity Rescale is to map point clouds with different physical scales and densities into a more comparable space for the encoder. More like we maintain the granularity we perceive the world, instead of the size.

So, it is not strictly necessary to always normalize the point cloud to a unit sphere such as [-1, 1] before encoding with Utonia. In our demo scripts, we use different rescaling strategies for different domains: object point clouds are typically rescaled to [-1, 1], while indoor scenes are usually scaled with a factor of 0.5, and outdoor LiDARs are scaled with a factor of 0.2.

This parameter can be adjusted depending on the granularity you want the model to perceive and your data customs. The main thing to keep in mind is that, after rescaling, the spatial range covered by the local attention windows should remain within the effective RoPE capacity (the RoPE base).

[L-Reichardt]

Where these parameters empirically determined or is there an overarching "formula" on what scale to choose, e.g. depending on sensor range?

[Yujia-Zhang0913]

Hi, thanks for the follow-up.
These values are mainly chosen empirically. A practical starting point is the typical grid/voxel size used for each domain, since it already reflects the usual point density and geometric granularity of that data. And then rescale these to a standard granularity.

In other words, the goal is not to normalize all scenes to the same absolute size, but to map them into a comparable perceptual granularity or density for the encoder.

In practice, we choose the rescaling factor by:

1. After rescaling, the spatial span covered by each local attention window stays within the effective RoPE range.
2. The voxelization/convolution stage still preserves local connectivity for sparse convolution, i.e., points should not become overly sparse or isolated after discretization.

So, sensor range only affects the choice indirectly. What matters more is the combination of scene scale, point density, voxel size, and the perceptual granularity you want the encoder to operate on.