[Talk-GB] Building with mapseg

[TimSC]

On 30/05/10 08:50, James Hogan wrote:

> It seems to have most trouble where there are bordering lines, such as
> public buildings (york district hospital it really struggled with), or
> long terraces next to roads which it seems to want to rotate quite a
> bit.
Special buildings with dark borders have a weird desaturated band 
between the central area and the dark band (you have to look closely). 
This makes the transition in RGB colour sharper than other buildings. 
This is probably causing the really jagged edges in the output. I am not 
sure how to approach this but I'll have a think.

Long terraces I will look into. If the road overlaps the building, weird 
things will happen of course.

> It did also trace around the letters of text overlapping
> buildings without flagging up ("Foss Islands" just east of York), but
> I bet that's pretty difficult to detect and is probably a complement
> to the shape detection that it did so accurately.
>    
Yes, it is possible the building happens to shape around the map label. 
Because of the ambiguity, I don't think there is a simple way to detect 
this in all cases. It is simpler to manually check the result than to 
try to solve that!

> Since it takes so long to process (took me 1h50m on laptop with an
> intel at 1.8GHz), is there a single place somewhere we can upload the
> resulting osm files of interesting areas (this one's 6.5MB)?
>    
It would be efficient to save recalculating the same tiles repeatedly. I 
am not sure why we want to recalculate tiles at all? I expect some 
people will do local imports but do we need to share the intermediate 
polygons? I don't know at this stage but it's worth consideration.

Another approach would be to optimize the code. The program was 
certainly not written to be computationally efficient. I would expect if 
some fuctions were moved to use C++ (using boost-python or similar), it 
would go like a bat out of hell. Is there any interest in that? (I 
probably won't be at my dev computer for about a week, if I get around 
to try this.)

> On a side note, has anybody noticed a consistent tendency for existing
> independently surveyed roads to be offset northwards (by around 5-10
> metres) from the OS data (vectormap and streetview)?
I am not seeing any problems personally in SE england. (I have not 
looked at vector map). I can't think of any cause, except if there was a 
problem with the transform used. What editor are you using? Do you see 
the offset if you read a POI map reference of the original image, 
convert the grid ref using OSTN02 and compare to OSM?

Tim