Questions about in_channels=6

[Quest4Science]

I noticed that in the released code, the model is initialized with in_channels=6. Does this mean that the xyz coordinates are used twice, or is there another reason for this configuration? I was initially expecting the inputs to be xyz coordinates plus normals (or possibly RGB).

Could you clarify:

During training, were the inputs strictly xyz coordinates only, or were normals (or any other features) also used?

If normals and/or RGB were considered, what was the reason to exclude them from the final input?

Do you foresee potential improvements if color information were incorporated as part of the sampled input?

This would help me better understand the design choices behind the input representation and its effect on the model’s generalization.

Thank you for your time and for sharing this impactful work!

[Colin97]

The in_channels = 6 is because we concatenate the point coordinates (xyz) with the point features (which are also xyz in this case). We have not experimented with RGB or normal features. We did consider using 2D renderings to extract high-dimensional 2D features and then unproject them back to 3D to serve as point features. However, we found that this approach significantly increases the computational cost while providing only slight improvements. Therefore, we decided to use only xyz features in our final version.

[Quest4Science]

您好，来信已收到，我会尽快回复。祝您天天开心 (*^▽^*)