Time-Lapse Mining from Internet Photos | Hacker News Reader

Time-Lapse Mining from Internet Photos

grail.cs.washington.edu

408 pointsdoh36 comments

This is an interesting counter-argument to the question "why bother taking photos of the well-known place X, there are already thousands of (better) pictures of it?"

That's really interesting that the Wall Street bull moves over the years. I wonder if this is due to people leaning against it, or something like cleaning, or some other reason.

Was about to ask the same question. I was thinking the ground shifted over the years or something.

The bull actually appears to move, to have rotated.

Look at the floor of that segment and you'll notice a cobble stone being revealed over time that was definitely covered in the earlier photos.

That one is in the "failure cases" segment

The blurring is the failure case, not the movement of the statue

Very interesting project. They show a few "failure cases" at the end of the video; I'd be interested to see a few more of those. The ones they showed really didn't look all that bad.

I wonder what the copyright on such a timelapse would look like?

Some of the photos they used are CC licensed, so it would probably be CC-BY-SA.

My problem would the images that are copyrighted, unless they specifically pulled only CC images, which would limit photo numbers by quite a bit.

I was thinking the exact same thing - the failure cases are possibly more interesting than the 'successful' ones. The Wan Chai skyline particularly seemed to pulse between day and night in a very interesting manner.

This is great work that seems immediately useful to areas like conservation and art.

The images were heavily homoginised which was good for continuity but perhaps obscures details.

Using the timelapses as a scrubbable index to the original images would be a cool tool.

The images probably needed this homogenization because they don't match 100%, being taken from slightly different places and angles (so you not only get different framing but also parallax issues).

This reminds me of https://vimeo.com/63653873, where they used Google Maps Street View to make hyperlapses.

The video displays a world map with the locations where they found enough photos to make time lapses. Does anyone has a link to this map? It would be great to discover the best time of the year to visit some places, like Salto Angel, the waterfall in Venezuela, or Lençóis, in Maranhão State - Brazil.

Imagine combining this with Photosynth or whatever the project was that could stitch random photos into a 3D scene.

Was thinking that back than when Microsoft could have acquired Yahoo, that this would eventually lead to a quite comprehensive 4-dimensional point cloud of everything ever photographed.

Or ever existed. Time travel anyone?

Or determining the approximate time and location of unlabeled or uncataloged photos from all kinds of archives.

Oh, the possibilities.

(Oh, the privacy nightmares.)

Creepy, and worth a novel.

A lot of the underlying machinery is similar: they both rely on "structure from motion" (sfm) techniques to automatically estimate both camera locations and 3d geometry of the scene simultaneously. And both works come from the same lab: the GRAIL group at the University of Washington.

(I postdoced in that lab for 3 years and the authors of this paper are friends/former colleagues.)

Cool. SFM is a really fascinating concept to me. How much of that work is commercially available to consumers?

More than people realize... Google and apple, for example, use sfm heavily to compute their 3d maps (in Google earth and the apple equivalent). Google also uses it in various other products.

Very very cool, but I was expecting to see a time lapse of a mining operation and it's effects. Still want to see that!

have you seen manufactured landscapes?

http://en.wikipedia.org/wiki/Manufactured_Landscapes

https://www.youtube.com/watch?v=o8cpCL8FSFM

beautiful film, there is one scene where the camera is tight on a 2 meter diameter tire dump truck and then slowly pans out until the dump truck is only the size of a pixel revealing a mining operation that fills the whole scene

Thanks for the suggestion!

A more specific title might be: "Internet Crowd-sourced Timelapse Integrator"

With all such results there is one question: are they releasing code? Otherwise... it's a nice demo to watch, but impossible to use (and even: hard to increment on it).

Well that'd be research code anyway :) You'd probably better off rebuilding from scratch based on the research paper. The software shouldn't be that complicated in the end, the main challenge (to me) is gathering that 86M images dataset.

Research code is infinitely better than no code. Research papers are almost never complete enough to reimplement it (seriously - I met a lot of academicians doing it, and most of the time it was not possible without contacting authors; just - text does not compile).

Vide:

http://academia.stackexchange.com/questions/23237/why-are-pa...

http://academia.stackexchange.com/questions/37370/should-i-s...

(A similar things holds for data as well.)

> Upcoming

> Code and more results will be published soon.

In retrospect this is a very obvious idea. Brilliantly done!

Very impressive technology. This is just the beginning of leveraging the massive amounts of data out there in innovative new ways.

I thought (inspired by Person of Interest TV show and photo supersaturation) after the Boston bombing it would be a good project to use publicly available photos, video, security camera video to track or search for suspects.

Or it could be used to search for a kid such as during an Amber Alert.

But I have no skills to do this :(

There's a lot of news and 'new' stuff on the internet. And then occasionally you see something like this which is truly a fresh idea. I feel inspired when that happens!

This is amazing! Would love to see more, and longer, timelapses!

This shows how the natural world around us is slowly, but always changing. It never occurred to me that waterfalls changed like this!

Reply on news.ycombinator.com