[OSM-talk] (magical?) road detector

Ido Omer Ido.Omer at microsoft.com
Sun Feb 6 19:32:01 GMT 2011


Hi Thibault,

Thanks again for your response.
I might be a bit overoptimistic, but I think the tool will be able to handle most cases as long as you use it properly.
We have preliminary nice results (internally) even in urban area and places with shadows etc. as long as you are "detecting" a reasonable road.
We don't believe the attached examples are typical examples. They are nice for demonstrating the limitations of the algorithm and computer vision techniques in general, but we don't think the average user will try to model a way that spans over a few roads with many junctions. In the examples you sent, it is not clear which is the road that connects the two points in question, since there are many valid possibilities between them.
We plan on having an exploration mode that will suggest many roads in the region; however, it will be based on a different module and will use different techniques.
I posted a (partial) list of the road detection algorithm's limitations. I am confident that once you apply them, you will find performance has improved.
There are also several modules that we have in house, and have already tested, but still have not incorporated in the "road detection pipeline", for instance, car detector that can help a lot in dense urban areas.
As I stated earlier, Bing is not putting its full weight on trying to solve this problem, and our (development) resources are quite limited, but we at Bing believe that our effort will create value for the OSM community and help build better map editing tools.

Regards,
Ido

-----Original Message-----
From: Thibault North [mailto:tnorth at fedoraproject.org] 
Sent: Saturday, February 05, 2011 6:02 PM
To: talk at openstreetmap.org
Subject: Re: [OSM-talk] (magical?) road detector

A few remarks again below:

On February 5, 2011 06:57:42 pm Ido Omer wrote:
> Hi Elizabeth,
> 
> Like in any interesting problem, we knew we will not be able to come 
> up with a general solution for all roads under any condition. We 
> assume most people are mapping asphalt/concrete roads and we decided 
> to focus our efforts there.

And even in that case, good luck! :-)
See for example these two shots from Bing maps, Montreal, and how bad is the road extraction:
http://tnorth.ch/blog/public/Dijkstra/tests/montreal1_out.jpg
http://tnorth.ch/blog/public/Dijkstra/tests/montreal2_out.jpg
(Showing, by the way, that a shortest-path algorithm is not always what we want to extract data in a robust way...)

Image processing in that area is a pain: shadows, trees over the road, gray roofs that look pretty much like a road...

> Eve there our solution is far from being perfect, but we believe in 
> most circumstances it has a value that is larger than zero.

I don't really know what to think about that. The tool will maybe never be efficient to handle most cases, but can perhaps sometimes help, and it is a sufficient reason to try and improve it.
Allowing the user to have some kind of control might help the detection to be done properly.

> Strictly speaking, the current version doesn't look for road patterns; 
> the next version might do that and provide a more general solution. It 
> will be useful for us to learn what the community needs are; I'll be 
> happy if you could refer me to a few examples.

In the mapping process (with JOSM or such tool), following roads is not really a problem, especially when they are not too sinuous (and that's when the road detector works well...). It can be done in a few clicks. Maybe the tool should try and act differently (but that is more GUI/UI related), and we could imagine the following scenario:
- The user wants to map roads and selects a "road extraction" tool.
- He roughly follows a road, maintaining a click (as you would do to paint with a brush in image processing softwares)
- The algorithm knows the approximate path, and tries to fit exactly the center of the road.

Regards,
Thibault

> BTW, by not using the detector its usefulness should be around zero, 
> it is hard for me to imagine how can it drop much lower than that (at 
> least for the long run, after you wasted a couple of minutes and found 
> out it does not serve your needs)
> 
> Thanks,
> Ido
> 
> -----Original Message-----
> From: Elizabeth Dodd [mailto:edodd at billiau.net]
> Sent: Saturday, February 05, 2011 2:27 PM
> To: talk at openstreetmap.org
> Subject: Re: [OSM-talk] (magical?) road detector
> 
> I've made one attempt only at tracing a dirt road in dry country with 
> the detector. I found its usefulness less than zero, as the system 
> told me that too many tiles were involved and quit. Zooming in is not 
> always practical to spot these roads, where the pattern recognition is 
> a very long straight feature on a photograph, and same colour as the 
> surrounds for a dirt road in dry country, and a dark colour for a 
> railway. Having spotted the road then it is easier to find in zoomed 
> images when looking for curved bits through (dry) waterways. I went 
> back to doing it by hand, so for me it had a uselessness of less than zero.
> 
> _______________________________________________
> talk mailing list
> talk at openstreetmap.org
> http://lists.openstreetmap.org/listinfo/talk
> 
> 
> _______________________________________________
> talk mailing list
> talk at openstreetmap.org
> http://lists.openstreetmap.org/listinfo/talk

_______________________________________________
talk mailing list
talk at openstreetmap.org
http://lists.openstreetmap.org/listinfo/talk




More information about the talk mailing list