[postgis-devel] I want boxes

[Paul Ramsey]

Just to point out:

The problem with the centroid is that, for large long features, lt's
overselective. So to get the "full potential set" of things you'll
have to winnow down, you need a large first limit.

The problem with the box is that, for large long features, it's
underselective. So you'll get lots of things that aren't actually very
close still showing up al their boxes interact with the candidate,
causing zero-distance returns. The net result is that you'll still
need a large first limit, to ensure that in addition to all the
zero-distance false positives you also have the non-zero results that
might actually be the closest.

Which is not to say the box-distance might not be better in the end,
just I'm finding in hard to visualize an extreme improvement.

P

On Tue, Sep 27, 2011 at 9:57 AM, Paragon Corporation <lr at pcorp.us> wrote:
> If it weren't for the fact it's the centroid of the box, I would agree with
> you.
>
> But honestly in many of my cases the centroid of my object is no where near
> the centroid of the box so pretty useless
> to base anything on it.  Works fine for parcels but not much else I have.
>
> for the case of polygons etc.  as Nicklas and I think others have pointed
> out, the objective is a little different
>
> not to get the best sort but faster compute the maximum distance you need to
> stretch out to grab the top 10.
>
> Though I'm curious how Oracle's NN which is also imperfect works.  Haven't
> paid much attention to that.
>
> Thanks,
> Regina
>
>
>
>> -----Original Message-----
>> From: Paul Ramsey [mailto:pramsey at opengeo.org]
>> Sent: Tuesday, September 27, 2011 12:52 PM
>> To: Paragon Corporation
>> Cc: Sandro Santilli; nicklas.aven at jordogskog.no
>> Subject: Re: I want boxes
>>
>> I can almost guarantee you that boxes will be just as bad but
>> in a different way... and centroids are always relative to
>> the box :) I will think on doing box distance.
>> P.
>>
>> On Tue, Sep 27, 2011 at 9:48 AM, Paragon Corporation
>> <lr at pcorp.us> wrote:
>> > As I pointed out to strk -- the centroid appears to be ceontroid of
>> > the box which is even less useful than I had originally
>> thought.  That
>> > said box distance is way better.
>> >
>> > Is it possible to have the distance be box based instead of
>> centroid.
>> > Centroid isn't terrible useful.
>>
>
>
>