back to article UK's map maker Ordnance Survey plays with robo roof detector

Britain's map maker is demonstrating that machine learning isn't all hot air – but has discovered just how much donkey work is involved. The Ordnance Survey set up an ML experiment to identify roofs from its remote sensing data (satellite and aerial imagery), and found the trial reached 87 per cent accuracy within a week, from …

  1. tiggity Silver badge


    "The results prove that automatically identifying changes to certain types of Land Parcel Boundaries – especially, drains/ditches/dykes, walls, or fences (which are smaller than the spatial resolution of the input imagery data) – is very difficult"

    And in those scenarios a PITA for people too, refining algorithms in those cases is still not going to get "decent" results (as in better than human and in high 90s accuracy level), only option is higher resolution imagery.

    1. SkippyBing

      Re: Duh

      Yes, I did think 'what were they expecting' with that, it's like asking someone to spot the invisible man.

      1. Yet Another Anonymous coward Silver badge

        Re: Duh

        Getting computers to extract and measure features smaller than the resolution element, or below the noise floor is what we have been doing with them for 40years. Of course smaller than a pixel AND below the noise floor is trickier but with good data it is easy.

        We are currently taking medical images (CT and MRI) with only 1/10 the resolution and processing it to get the same image as the full res scan - but with a fraction of the patient scan time and dose.

      2. Adam 52 Silver badge

        Re: Duh

        The OS guys who do this are absolutely awesome at it. They really can tell that a jumble of blurred pixels is a fence and fence post. And, even more impressively, whether it's going uphill or downhill (it's stereoscopic imagery).

        1. John Smith 19 Gold badge

          "(it's stereoscopic imagery)."

          That I suspect makes a very big difference between what is and is not possible.

          Likewise 87% may not be bad if it lets humans focus on the doubtful images, one way or the other, eliminating the drudge work.

          It's not clear if this thing does that.

        2. Alan Brown Silver badge

          Re: Duh

          "They really can tell that a jumble of blurred pixels is a fence and fence post."

          Because they've seen it hundreds of times _AND_ been taught by someone what it was _AND_ seen it at various resolutions, so knwo what to expect.

          ML doesn't work like that. You can't show it the same image at varying resolutions and have it learn "this is what X looks like when we vary the view"

          1. Nick 6

            Re: Duh

            Yes, yes you can through data augmentation techniques. Its quite a common requirement to be able to train for detection at different distances, resolution, pose angles, skew etc.

  2. Swiss Anton

    Ho ho ho

    Maybe Santa could make use of a refinement of this to determine whether a roof has a chimney.

    It could help to optimise things on Christmas Eve.

    1. Alister

      Re: Ho ho ho

      However, determining if there's a gas or electric fire at the bottom of the chimney is still beyond AI.

      Yes children, that muffled thumping noise and swearing is just Santa trying to deliver the presents...

  3. frank ly

    The Next Stage

    Farmers start applying camouflage patterns to the roofs of the new barns they built in a convenient hollow in their land.

    1. Nick Ryan Silver badge

      Re: The Next Stage

      Plant trees to break up the hard lines and shadows and apply camouflage pattern. That would probably do the trick.

    2. Pen-y-gors

      Re: The Next Stage

      Do it properly - put a living roof on the barn - grass, herbs, flowers, shrubs. That would really confuse them.

    3. Anonymous Coward
      Anonymous Coward

      Re: The Next Stage

      You could also fool it by planting different coloured plants in a house formation or painting the ground.

      1. Jason Bloomberg Silver badge
        Thumb Up

        Re: The Next Stage

        Turf the roof and plant a hedgerow above the guttering. Not only will it not be detected for what it is but an automatically generated cheque for the additional hedgerow should soon arrive in the post.

    4. Anonymous Coward
      Anonymous Coward

      Re: The Next Stage

      Something like this Muppet?

      1. Ken Moorhouse Silver badge

        Re: Salfords Castle

        Did he realise that if he kept it he might have been slapped with a crenellation licence tax?

  4. Paul Hovnanian Silver badge

    Just a sec.

    Barbra Streisand is on the other line.

  5. Anonymous Coward
    Anonymous Coward

    Hundreds of years...

    It's taken hundreds of years to get the maps to where they are now so why the rush all of a sudden to get rid of more jobs? I can't see anyone really benefiting even if they can detect the type of roof that much quicker.

    Bit by bit we appear to be putting a bigger tax burden on fewer people as quick as we can.

    If AI does as well as hoped just who will have a job?

    I'm all for other people working to pay my benefits but rather scared that no one will be left soon :)

    1. Dave 126

      Re: Hundreds of years...

      For sure, the role of ML in reducing the number jobs is a valid topic, but one that perhaps is better addressed directly and not on every article - after all, if computers didn't make tasks quicker and easier, why bother with them? In any case, the jobs issue goes way back, before such things as combine harvesters and Spinning Jennies.

      Whilst we have been using maps for hundreds of years, we are now using them for far more than navigation and settling land disputes.

      1. Anonymous Coward
        Anonymous Coward

        Re: Hundreds of years...

        "For sure, the role of ML in reducing the number jobs is a valid topic, but one that perhaps is better addressed directly and not on every article "

        ...fair point :)

  6. David 132 Silver badge
    1. Bronek Kozicki

      Re: OK, cleverclogs ML algorithm

      Flat, obviously.

    2. Paul Hovnanian Silver badge

      Re: OK, cleverclogs ML algorithm

      How about this one?

  7. Alan J. Wylie

    In about 1980

    when I was in Southampton installing a very early automatic line following map digitising system, I remember being told that aerial photos weren't sufficiently accurate - it's the house walls that go on the map, not any overhanging eaves.

    For more info on the Laser-Scan Fastrak (which had a display resolution of 14000 by 10000)

  8. ThaumaTechnician

    Automated land/house value assessment process?

    Been (t)here, done that. Didn't work so well.

    "In [one] case, the Service New Brunswick [software] concocted a figure of $79,780 in renovations to her tiny 860 square foot home — or "major improvement change" in the parlance of the department — and then raised her tax bill a stunning 52 per cent because of the phantom improvements."

  9. Anonymous Coward
    Anonymous Coward

    land parcels and building schematics are already captured on creation and alteration

    Last time I had anything to do with this the local councils already had this data so the only need for this algorythm is to identify buildings without planning permission.

    As with the reports of building within barns etc, which are intented to prevent aerial and street level viewing then this algorythm adds nothing, the ones that built out of view were actually caught before the deadline were from anonymous tipoffs, and very few from someone spotting unsanctioned building on an aerial view.

    It's a nice tool but unless the idea is to remove local council workers then it doesn't have an application.

    We already have enough people with degree level colouring-in unemployed do we really need more?

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Other stories you might like