Can you explain how you use the vantage tree? What are you storing/searching for? What error metric are you using for matching?

I think you briefly talked about using heading change and distance for encoding your polyline. So if I understand correctly, something like:

[forwardDistance1, angleDelta1, forwardDistance2, angleDelta2, ...]

Then are you using it kinda like a kd-tree (filter everything within a threshold of forwardDistance1, then angleDelta1, then next dimension and so on)? But then how do you handle the initial scaling? What about cumulative error in heading or distance?

Congrats on the amazing work by the way!

thanks! I am using a metric from the dollar gesture recognizer

http://depts.washington.edu/madlab/proj/dollar/index.html

which gives a value for how close to polylines are (it normalizes them and does some distance calculations)

the way the vp tree works is you provide a set of data and a metric for distance (in my case, the data was the polylines and the metric was from dollar) and it computes the structure. As long as the metric observes some basic principles (I think it's called triangle inequality) the spatial division will work and you can do a fast search for nearest neighbors.

This is extremely inspiring for me. Thank you for sharing how you did it!

thanks! that's great to hear