A Little Question in Studying the Paper

[emmalzw]

Great job! I have a question while studying your literature. The original text mentions : ... Collecting boundary points from the input point cloud is straightforward with the ground truth label. However, after sub-sampling, it is difficult to obtain a proper definition of boundary point set following Eq. (1), due to the undefined label for sub-sampled points [14]. .... In the original point cloud scene, each point has a idex number. If during the downsampling process, each point has a subsampling idex. For example, with 1 million point clouds, we firstly downsample 40960 of them with the subsampling idexs in the original point cloud. We can find the truth label of the original point cloud through the subsampling idex, and thus obtain the boundary points. Is this also possible?

[LiyaoTang]

Hi @emmalzw ,

Thanks for your interest.

Yes, for the case you mention, we can directly obtain it in the way that you suggest.

However, for a more general case, where we might use gird-sampling as in KPConv, there is no 1-to-1 correspondence between the original and sub-sampled points.

Best, Liyao