One Step


[via engadget, physorg, etc..]

For those of you who remember my post from three years ago, we asserted that the best way to deal with images of people, cars, and other moving objects in Street View (and by extension Microsoft’s own StreetSide product) was not to blur, but to remove them entirely. They could always be re-added in a variety of forms, ideally as dynamic 3D objects.

Looks like it’s taken a while, but the simplest method of removing people — comparing successive images for similarities — works pretty well. Congrats to the grad student who pulled this off. Nice work.

  1. #1 by Ron Schott on August 8, 2010 - 10:45 am

    I guess dogs don’t deserve the same censorship that humans do.

  2. #2 by avi on August 8, 2010 - 11:17 am

    For dogs, I think some censorship might be appropriate here, especially when they ignominiously lick their own privates. But hey, a dog’s got to do what a dog’s got to do.

    For people, yes, it’s much more complicated. When dogs begin to sue, vote, or become the target of paparazzi, then we’ll talk.

  3. #3 by Anonymous on August 9, 2010 - 12:26 am

    You sound a bit condescending. What’s up with that?

    • #4 by avi on August 9, 2010 - 8:26 am

      Condescending towards dogs? If you want to critique my writing style, you’ll need to be more specific.

  4. #5 by Ira on August 10, 2010 - 1:43 pm

    Looking at the images on the link you provided, it appears this grad student uses only a single original image, identifies people automatically (or perhaps manually) and then copies over groups of nearby pixels to obscure the people (apparently manually). One sample on the site you linked to shows a woman disappear but most of her umbrella remains. A single image obviously cannot be used to accurately reconstruct whatever was really behind that person.

    Your idea was “oversampling” – multiple images from the same viewpoint during successive time periods. The images would then be compared, pixel-for pixel, and use majority-rule to pick the one to be used in the final image. An advantage of your method is that it could be totally automatic and would remove not only people, but also dogs (and even cats) but leave things like statues and cigar store native Americans that might be blurred or erased by automated people-removal software.

    A disadvantage is the need for multiple images spaced out over the time it might take for a person to move out of the way. Perhaps if images were taken from a moving auto at a higher rate on a single pass, the successive images, from slightly different viewpoints, could be processed to construct a 3D model, and then processed further to identify objects, like street lamps, that do not move relative to other fixed things, and people, who do and who could thus be eliminated or at least have their facial features blurred, all automatically.

(will not be published)