Add header only cuspatial::quadtree_on_points

[trxcllnt]

Closes #566.

[harrism]

Good work so far, this is a big task. I didn't do a complete review. I will re-review after you go over these comments.

[harrism]

This note is confusing because it seems to contradict below where it says "@param min_size Minimum number of points for a non-leaf quadtree node." If it's the minimum, then how come the non-leaf nodes "should" have fewer than min_size points? Either the parameter is misnamed or the docs need rewritten?

[trxcllnt]

You're right -- reading through the code, I think this is the max number of points allowed in a node before the node is subdivided into 4 new leaves.

[zhangjianting]

@harrism @trxcllnt It is the "min_size" to qualify as a non-leaf node, which should be the same as the "max_size" as a leaf-node. Just to want to add that the numbers of nodes for the last level nodes (which need to be leaf nodes) is not limited by the threshold. When min_size/and or max_depth is small, many (could be most) last-level nodes will have large numbers of points under them.

[zhangjianting]

To answer Mark's question directly, non-leaf nodes need to be split and non-last-level nodes after the split should have fewer nodes than min_size. I agree changing it to "max_size" to indicate maximum number of non-last-level leaf nodes might be easier to parse.

[harrism]

Just a couple more things.

[harrism]

"applied" doesn't reflect that those differences are divided by scale... If there is an intuitive explanation, let's add that. And if not, then put the exact math in here. See #397

[harrism]

As discussed, should probably use a reciprocal of the current scale to make the scale factor actually a directly proportional scale factor (which would be more intuitive).

[harrism]

Further discussion: let's save any semantic / breaking change for a later PR if necessary.

[isVoid]

Sorry about the late review. I delved into the reference in depth but still got lost in the remove_unqualified_quadtree (and Phase 2 in general, probably running some numbers will help). According to this comment: #639 (comment). It feels like max_size still should be called min_size, otherwise it makes remove_unqualified_quads strange to read. We should give it a more verbose name e.g. min_non_leaf_node_size.

@zhangjianting Is it possible to have a leaf node that's not on the last-level? aka is quadtree full?

I find it fascinating to compute the proximity between points using dilated integer and cluster them simply by dividing the Morton code by 4. Apparently scale is an important factor here. How to pick scale?

Not a lot of code smell in terms of the header only refactoring except for a few small issue (IIFE and structure binding, not all are labeled).

[harrism]

@gpuccibot merge

[harrism]

@gpucibot merge