Kd Tree Revisited

After giving it a bit of thought, I’ve found a way to simplify the nearest neighbour search (i.e: the closest method) for the KdTree I implemented in my previous post.

The improvement

That post implemented the nearest neighbour search by keeping track of the tree’s boundaries (through AABB), and each of its sub-trees (through AABB.split), and testing for the early exit condition by computing the distance of the search’s origin to each sub-tree’s boundaries.

Instead of explicitly keeping track of each sub-tree’s boundaries, we can implicitly compute it when recursing down the tree.

To check for the distance between the queried point and the splitting plane of inner nodes: we simply need to project the origin onto that plane, thus giving us a minimal bound on the distance of the points stored on the other side.

This can be easily computed from the axis and mid values which are stored in the inner nodes: to project the node on the plane we simply replace its coordinate for this axis by mid.

Simplified search

With that out of the way, let’s now see how closest can be implemented without needing to track the tree’s AABB at the root:

# Wrapper type for closest points, ordered by `distance`
@dataclasses.dataclass(order=True)
class ClosestPoint[T](NamedTuple):
    point: Point = field(compare=False)
    value: T = field(compare=False)
    distance: float

class KdTree[T]:
    def closest(self, point: Point, n: int = 1) -> list[ClosestPoint[T]]:
        assert n > 0
        res = MaxHeap()
        # Instead of passing an `AABB`, we give an initial projection point,
        # the query origin itself (since we haven't visited any split node yet)
        self._root.closest(point, res, n, point)
        return sorted(res)

class KdNode[T]:
    def closest(
        self,
        point: Point,
        out: MaxHeap[ClosestPoint[T]],
        n: int,
        projection: Point,
    ) -> None:
        # Same implementation
        self.inner.closest(point, out, n, bounds)

class KdLeafNode[T]:
    def closest(
        self,
        point: Point,
        out: MaxHeap[ClosestPoint[T]],
        n: int,
        projection: Point,
    ) -> None:
        # Same implementation
        for p, val in self.points.items():
            item = ClosestPoint(p, val, dist(p, point))
            if len(out) < n:
                out.push(item)
            elif out.peek().distance > item.distance:
                out.pushpop(item)

class KdSplitNode[T]:
    def closest(
        self,
        point: Point,
        out: list[ClosestPoint[T]],
        n: int,
        projection: Point,
    ) -> None:
        index = self._index(point)
        self.children[index].closest(point, out, n, projection)
        # Project onto the splitting plane, for a minimum distance to its points
        projection = projection.replace(self.axis, self.mid)
        # If we're at capacity and can't possibly find any closer points, exit
        if len(out) == n and dist(point, projection) > out.peek().distance:
            return
        # Otherwise recurse on the other side to check for nearer neighbours
        self.children[1 - index].closest(point, out, n, projection)

As you can see, the main difference is in KdSplitNode’s implementation, where we can quickly compute the minimum distance between the search’s origin and all potential points in that subspace.

2024-08-17

https://belanyi.fr/2024/08/17/kd-tree-revisited/ Bruno BELANYI

Programming