don't break when there are few or collinear points #62
Conversation
fixes #50 note: I had to disable the tests for outeredges, because I'm not using the same outeredges (but they do point to the same nodes).
|
Rebased (with some clean-ups) on the indexHull branch. Updated the test at https://observablehq.com/d/5e4da35146afd543 |
|
Personally this looks too hacky to be worth the benefit in this form... Can we approach the issue some other way, including changes to Delaunator upstream? The "try/catch" just feels like a blanket cover fix, and depending on a constant number to designate "far away" doesn't feel right too — e.g. I can easily imagine Delaunay being fed some big numbers in this range. |
|
Agreed. In d3-geo-voronoi in particular, the numbers can be very large after the stereographic projection — and there is already too much numeric instability (see Fil/d3-geo-voronoi#10). I wonder if the solution wouldn't be to return an empty With this approach Delaunator would not lie about triangles ("here's your 0 triangles"), and just a little bit about the hull. d3-delaunay could work with that pseudo-hull and create the list of infinite cells in that order. |
|
New code, much cleaner. Updated the tests at https://observablehq.com/d/5e4da35146afd543. Needs this change to Delaunator: https://github.com/Fil/delaunator/commit/b96802021867a788c2c2111085562a086996b52c |
the strategy is the following: - instead of throwing an error for collinear input, Delaunator returns a degenerate structure with hull = the valid points, sorted along an arbitrary axis (https://github.com/Fil/delaunator/commit/b96802021867a788c2c2111085562a086996b52c) - the cases with 1 or 2 valid points are handled specifically; - for 2 points, craft fake triangles to create the correct vectors - the case with 3 or more valid point is handled by jittering the points (deterministically); however neighbors are handled specifically, using the ordered list returned as "hull" by Delaunator
with a possibly superfluous test on i0 to avoid infinite looping if no point is regular fixes #55
… less than 2 points (ie no triangulation)
|
Superseded by #64. |
* return an ordered array of points if they are collinear (see d3/d3-delaunay#62) * test * simplify: collinear inputs can be sorted by x (or by y if all x are identical)
fixes #19 and #20
gross (try/catch)this is not ready to merge: there are issues with (at least):naming (this.__property is a placeholder)tuning the value ofg(1e9 works in the examples but creates some instability, as evidenced by the fact that lines are not neat — with 1e8 they are perfectly in the middle).should the API allow to request a strict Delaunay — or on the contrary, offer this "robust" approach as an option?