[OSM-dev] [GSoC] Video based speed limit detector

Kai Krueger kakrueger at gmail.com
Wed Mar 30 22:09:12 BST 2011


I had a short look at the proposal. Looking at some of the sample 
project proposals on the GSoC page, it looks comparable in detail.

Some thoughts of things you could consider mentioning in the proposal.

- You might want to mention a thing about data collection. For one you 
will need some sample video to play around with and test how well things 
work. More importantly though, you will likely need data to train the 
detection and recognition algorithms on. As you will likely use some 
form of supervised learning algorithm for those, you will need a 
training set of labeled examples, which you will need to create. I'd 
hope that the community will help collecting some raw video material 
that can be used, but it will then need to be labeled. Alternatively, or 
in addition, there might be some standard training sets for street signs 
and speed limits, given that there appears to be a reasonable set of 
research literature on the topic.

- It might be good to mention if once you have some basics going, if you 
want to take to project either more in the direction of building 
something that is fully integrated into the OSM editing ecosystem, or if 
you would rather take it more in the research direction of improving the 
quality of the detection algorithm, for example using more rich queues 
that the video gives you over the single image processing.

In that respect, it might be good to be slightly more specific on the 
deliverables, although you want to make sure that you don't over promise 
and leave enough freedom for inevitable changes along the line.

- You could also add a section to the end with a more speculative part 
of various possibilities of how the project could be extended in the 
future. Even though you are likely not going to be able to implement 
(all of) it, it might show that you have a good understanding of the 
topics involved and enough ideas to independently adapt if some of the 
things don't work out.

You might also want to put the project related parts of the application 
on a public facing page to give more people a possibility to give 
suggestions and comments.

But those are just some thoughts and it should be a good fun project to 
work on.


On 03/28/2011 09:08 PM, Keshan Sodimana | කේෂාන් සෝදිමාන wrote:
> Hi all,
> As i mentioned above. i drafted a project proposal and submitted it to 
> the GSoC site. so it would be great if any potential mentor can review 
> my proposal.feedbacks are welcome. here's the link to my proposal : 
> http://www.google-melange.com/gsoc/proposal/review/google/gsoc2011/keshansanjaya/1#
> (i didn't make this proposal publicly visible)
> Thanks !!!
> - Keshan
> On Tue, Mar 22, 2011 at 11:49 PM, Stefan de Konink <stefan at konink.de 
> <mailto:stefan at konink.de>> wrote:
>     Hash: SHA512
>     Op 22-03-11 08:33, Kai Krueger schreef:
>     > I'll possibly be able to mentor such a project, although I know
>     little about
>     > the code of any of the editors, so I'd be less able to help on
>     that side of
>     > things.
>     Since I was the mentor of the last project, there is a great number of
>     test material available to even build a recognizer. Video segmentation
>     is step two, not the first step.
Yes, the first step approach (and potentially even sufficient step) 
would be to treat each video frame as an independent still image.

Do you still have the material and write up for last years project, as 
that project is clearly very relevant to this one and thus should be a 
good source of inspiration.

>     If someone isn't able to even find a
>     sign on a still image, it is even harder to do it on motion pictures.
Depending on what you mean by "harder", not necessarily. Video gives you 
a lot of redundancy that you don't have in a still image. And so you can 
potentially accept a much higher false positive rate on the individual 
frames, as you combine the various predictions on the frames to reach a 
higher confidence, and thus you potentially can get away with a weaker 
>     For Dutch signs, and most likely many international ones on Wikipedia
>     SVG images do exist showing signs in the highest details possible. So
>     first things first:
>      - sign is present (x,y,w,h)
>      - classify sign
>      - segment video
>      - enhance recognitionrate on multiple images
>      - pinpoint the location of the sign in 3D
>     Stefan
>     -----BEGIN PGP SIGNATURE-----
>     Version: GnuPG v2.0.16 (GNU/Linux)
>     Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org/
>     iEYEAREKAAYFAk2I6C8ACgkQYH1+F2Rqwn3u/wCggw+qJzPbUuR60IzOclFlz3f8
>     i7gAnRm2+D2cBPWT3+Fd2hKIKdKghJqS
>     =reZv
>     -----END PGP SIGNATURE-----
>     _______________________________________________
>     dev mailing list
>     dev at openstreetmap.org <mailto:dev at openstreetmap.org>
>     http://lists.openstreetmap.org/listinfo/dev

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstreetmap.org/pipermail/dev/attachments/20110330/c4859df5/attachment-0001.html>

More information about the dev mailing list