I actually got to use a k-d tree before, I was naïvely brute forcing tiles together according to their 2d positions, the tiles being part of a larger image, and it was perfect.

Would run out of memory on a 96gb machine before, after it could process the tiles as quick as they came in from the camera, only a couple held in memory at a time.

When you have the right data structure, man it’s like night and day performance!