7 ms·

I'm Google – chains of visually similar images and videos in Google (2011)

332 points by dhotson 11y ago

captn3m0 11y ago
Here's a statement from the creator: http://dinakelberman.com/imgoogle/imgoogle.html http://dinakelberman.com/imgoogle/imgoogle.html The relevant bit: >Firstly, lots of people ask if it's an algorithm or something. It's not! Just me searchin google. >Secondly, a lot of people assume this blog is therefore made predominantly by using the "visually similar" function on Google Image Search, which is a totally reasonable thing to assume. While I definitely employ that function in my attempts to search thoroughly (and love it for it's own beautiful results), surprisingly little of the piece is actually constructed using it! Visually Similar appears to employ a logarithm based mainly on color percentages in an image, and as I'm Google is based more often in conceptual similarities than color-wash similarities, my searching is almost entirely relient on keywords rather than searching by image. Of course, there are times when Visually Similar has helped with a transition or section here or there, but overall, it's not the way I work.
- danso 11y ago
  > Firstly, lots of people ask if it's an algorithm or something. It's not! Just me searchin google. Love that...I sometimes worry we're getting to a point where those who haven't programmed before don't realize they can do things without apps. Such as making a data visualization by drawing. Or math.
- TorKlingberg 11y ago
  I though the image changes felt too intelligent. Thank you for the clarification, Google Skynet is not upon us after all.
  - luke-stanley 11y ago
    I wouldn't go that far, have you not seen Google Deep Dream?
  - jxramos 11y ago
    Yes, the transition points were so incredible, and downright hilarious. Especially one with an explosion within a TV screen transitioning into actual sets of explosions. Makes a lot of sense now.
codewithcheese 11y ago
If Google deep dream was the visualizations you might see on LSD, Google chains is like following a conversation of people on LSD =D
- ucaetano 11y ago
  Fear and Loathing in Google Dreams: https://www.youtube.com/watch?v=oyxSerkkP4o https://www.youtube.com/watch?v=oyxSerkkP4o
juliann 11y ago
I can't find where the about link is, but here's the about text i found on the source code: I’m Google is an ongoing tumblr blog in which batches of images and videos that I cull from the internet are compiled into a long stream-of-consciousness. Both the searching and arranging processes are done manually. The batches move seamlessly from one subject to the next based on similarities in form, composition, color, and theme. This results visually in a colorful grid that slowly changes as the viewer scrolls through it. Images of houses being demolished transition into images of buildings on fire, to forest fires, to billowing smoke, to geysers, to bursting fire hydrants, to fire hoses, to spools of thread. The site is constantly updated week after week, batch by batch, sometimes in bursts, sometimes very slowly. <br><br> The blog came out of my natural tendency to spend long hours obsessing over Google Image searches, collecting photos I found beautiful and storing them by theme. Often the images that interest me are of industrial or municipal materials or everyday photo snapshots. I do not select images or videos that appear to be intentionally artistic. Happily, the process of researching various themes in this way has lead to unintentionally learning about topics I might never have otherwise, including structural drying, bale feeders, B2P, VAWTs, screw turbines, the cleveland pack, and powder coating. I feel that my experience wandering through Google Image Search and YouTube hunting for obscure information and encountering unexpected results is a very common one. My blog serves as a visual representation of this phenomenon. This ability to endlessly drift from one topic to the next is the inherently fascinating quality that makes the internet so amazing. A NOTE ON THE PROCESS Just wanted to add a note on how I make this blog, as I have seen people wonder the same couple things frequently. Firstly, lots of people ask if it's a algorithm or something. It's not! Just me searchin google. Secondly, a lot of people assume this blog is therefore made predominantly by using the "visually similar" function on Google Image Search, which is a totally reasonable thing to assume. While I definitely use that function in my attempts to search thoroughly (and love it for it's own beautiful results), surprisingly little of the piece is actually constructed using it. Visually Similar appears to employ an algorithm based mainly on color percentages in an image, and as I'm Google is based more often in conceptual similarities than color-wash similarities, my searching is almost entirely relient on keywords rather than searching by image. Of course, there are times when Visually Similar has helped with a transition or section here or there, but overall, it's not the way I work. I hope you enjoyed my first FAQ – Dina Kelberman
anc84 11y ago
Aww man, I thought this really was pure algorithms, not manual curation. It is still a wonderful project but my awe is gone. Would be great to see it as a real bot.
- spacemanmatt 11y ago
  Write it!
shrimplypibbles 11y ago
My thoughts exactly. What a waste of time
- jahnu 11y ago
  Urgh... talk about giving tech people a bad reputation as philistines. It's a wonderful, thoughtful project.
  - teagoat 11y ago
    It is a wonderful, thoughtful project, but it's an art project. While that's great, I don't come to Hacker News for the latest interesting art project. I think like many people that clicked the link and scrolled through the images, we made the assumption that these were put together using some algorithm that mined Google's 'visually similar' search and then somehow found the best results in order to create a set of images that 'flowed'. It's a little disappointing to learn otherwise, but it's now kicked off a train of thought about how I might automate something similar. :-)
    - dang 11y ago
      There's nothing wrong with interesting art projects on HN. Story diversity is the most important thing to preserve here.
- laferrera 11y ago
  This comment is a waste of time.
- dang 11y ago
  We detached this comment from https://news.ycombinator.com/item?id=10338685 https://news.ycombinator.com/item?id=10338685 and marked it off-topic.
arxpoetica 11y ago
Somebody needs to automate this.
- martin-adams 11y ago
  Like with Mechanical Turk right?
  - rcpt 11y ago
    Could end up like https://vimeo.com/19760997 https://vimeo.com/19760997
abrichr 11y ago
I predict that someone will make an algorithmic version of this and post it to HN within a month.
- hantusk 11y ago
  There is already an algorithmic version like this. Video here: https://vimeo.com/34949864 https://vimeo.com/34949864
  - abrichr 11y ago
    Interesting. It seems to alternate between being "stuck" in certain conceptual valleys, and jumping between different concepts quickly. Perhaps an effect more similar to the op could be achieved by enforcing some sort of smoothness and similarity constraints.
- dang 11y ago
  Please do!
jstanley 11y ago
How to automate this: - pick a seed image and then: - run one of those "caption generating" algorithms on it (are any open source? what are the best ones?) - feed the caption into google images - pick the first result - repeat Probably also wants something simple to prevent cycles and fixpoints.
jpalomaki 11y ago
"Back in the days" image morphing software was somewhat popular. It could be fun project to run these images through some automated tool to create a continuous video of one image morphing to another then to third and so on.
- martin-adams 11y ago
  You might not need to morph it, just play it back at 24 fps and the results might be quite interesting. Essentially the hard work of interpolating frames has been taken care of by choosing a slow progression of one image to the next.
dfar1 11y ago
The transitions are brilliant!
Dav3xor 11y ago
Some of the break rooms where I work have a computer you can do searches on. There are some Google image searches that just shine like diamonds. Most of them are somehow PBS related for some reason. "Bob Ross" "Fred Rogers" "Ernie and Bert", etc. I started an odd little game of people leaving google image searches on the break room machines. You'd walk in and.. "Hedgehogs!"
kraig911 11y ago
I wonder would happen if they took every n(20) as a start for another similarity pass.
shubhamjain 11y ago
I am oddly reminded of Pollard's rho algorithim[1], you start with a base image and only choose the first image that comes up in the result, in the end we will eventually return to an image that we have already traversed. I wish someone could find the base images that result in the shortest and longest cycles. [1]: https://upload.wikimedia.org/wikipedia/commons/4/47/Pollard_rho_cycle.jpg https://upload.wikimedia.org/wikipedia/commons/4/47/Pollard_...
- accounthere 11y ago
  My guess is the shortest would usually be the very next one since Google Image is good at finding the same image file in different websites. There would have to be a constrain for how close to the last image the next one is or you'd end up in a one image loop.
digitus 11y ago
We're working on something really similar to help people explore visual content. https://posters.dreamitget.it/ https://posters.dreamitget.it/ https://watches.dreamitget.it/ https://watches.dreamitget.it/
MrBra 11y ago
I'm assuming porn is filtered out. Given that porn represents the vast majority of Internet content, it would be curious to see how quickly a sexually explicit image would pop up, and how funny the connection with the previous non sexual image would be.
- MrBra 11y ago
  Oh, it's not an algorithm. Give me back my upvote :)
- eonw 11y ago
  i don't think porn constitutes the vast majority of internet content by any means, or even the majority at all.
  - MrBra 11y ago
    It seems that according to modern estimates this thing about porn being the greatest part of the Internet is no way close to reality and was more of just a legend that got told over and over and ended up being accepted for truth. But honestly I myself don't know and can't be bothered to investigate more. I just know there's enough of it :)
gloves 11y ago
This is cool! I love how on Google when I do an image search on myself I can go from Brad Pitt one minute, to an old man the next. That's not a complaint though. The technology is always improving and I live in constant wonder of the things people can create from nothing :)
Cerium 11y ago
Very cool project. As an interesting side effect I'm sure Dina learned a lot of obscure trade specific words while working on this.
luluganeta 11y ago
This work is from the same year, also really interesting: http://sebastianschmieg.com/searchbyimage/ http://sebastianschmieg.com/searchbyimage/
accounthere 11y ago
It would be nice if you could change the seed image. Someone needs to automate this.
danschumann 11y ago
Kevin Bacon
danschumann 11y ago
Kevin Bacon
hellbanner 11y ago
Reminds me of http://translationparty.com/ http://translationparty.com/ which converts between Japanese & English using google translations
hellbanner 11y ago
A few scrolls down there is "This video was removed by Youtube for privacy reasons".. does the related image algorithm still use what image was there, for calculating next image?
- justinator 11y ago
  There is no algorithm, this was done by hand.
davidhariri 11y ago
This is so cool!
socialmediageek 11y ago
This is cool. It could become a great business if there are links to buy those items. Like Amazon's similar items (which is still within Amazon) - this may help several businesses.
yazaddaruvala 11y ago
When I read the title "I'm Google - ..." a very small part of me was really, really excited that maybe, just maybe, some system within Google became self aware, learnt english, and wrote a blog declaring its identity.
- tajen 11y ago
  Man, you've made my day. I've always imagined the first PR about transient AI as "Dr. has applied for a patent for Y" or "How can Siri answer this, although she only knows that". But the idea of a system becoming transient and Google employees not knowing where it is stored inside their datacenters would be plausible. Assuming AI is possible. Hopefully the first thing this AI will declare is whether it stores data on its EU server.
cognivore 11y ago
I love the one that goes from cookie dough to car plowing through the sand dune!

332 points by dhotson 11y ago

Here's a statement from the creator: http://dinakelberman.com/imgoogle/imgoogle.html http://dinakelberman.com/imgoogle/imgoogle.html The relevant bit: >Firstly, lots of people ask if it's an algorithm or something. It's not! Just me searchin google. >Secondly, a lot of people assume this blog is therefore made predominantly by using the "visually similar" function on Google Image Search, which is a totally reasonable thing to assume. While I definitely employ that function in my attempts to search thoroughly (and love it for it's own beautiful results), surprisingly little of the piece is actually constructed using it! Visually Similar appears to employ a logarithm based mainly on color percentages in an image, and as I'm Google is based more often in conceptual similarities than color-wash similarities, my searching is almost entirely relient on keywords rather than searching by image. Of course, there are times when Visually Similar has helped with a transition or section here or there, but overall, it's not the way I work.

danso 11y ago
> Firstly, lots of people ask if it's an algorithm or something. It's not! Just me searchin google. Love that...I sometimes worry we're getting to a point where those who haven't programmed before don't realize they can do things without apps. Such as making a data visualization by drawing. Or math.
TorKlingberg 11y ago
I though the image changes felt too intelligent. Thank you for the clarification, Google Skynet is not upon us after all.
- luke-stanley 11y ago
  I wouldn't go that far, have you not seen Google Deep Dream?
- jxramos 11y ago
  Yes, the transition points were so incredible, and downright hilarious. Especially one with an explosion within a TV screen transitioning into actual sets of explosions. Makes a lot of sense now.

codewithcheese 11y ago

If Google deep dream was the visualizations you might see on LSD, Google chains is like following a conversation of people on LSD =D

ucaetano 11y ago
Fear and Loathing in Google Dreams: https://www.youtube.com/watch?v=oyxSerkkP4o https://www.youtube.com/watch?v=oyxSerkkP4o

juliann 11y ago

I can't find where the about link is, but here's the about text i found on the source code: I’m Google is an ongoing tumblr blog in which batches of images and videos that I cull from the internet are compiled into a long stream-of-consciousness. Both the searching and arranging processes are done manually. The batches move seamlessly from one subject to the next based on similarities in form, composition, color, and theme. This results visually in a colorful grid that slowly changes as the viewer scrolls through it. Images of houses being demolished transition into images of buildings on fire, to forest fires, to billowing smoke, to geysers, to bursting fire hydrants, to fire hoses, to spools of thread. The site is constantly updated week after week, batch by batch, sometimes in bursts, sometimes very slowly. <br><br> The blog came out of my natural tendency to spend long hours obsessing over Google Image searches, collecting photos I found beautiful and storing them by theme. Often the images that interest me are of industrial or municipal materials or everyday photo snapshots. I do not select images or videos that appear to be intentionally artistic. Happily, the process of researching various themes in this way has lead to unintentionally learning about topics I might never have otherwise, including structural drying, bale feeders, B2P, VAWTs, screw turbines, the cleveland pack, and powder coating. I feel that my experience wandering through Google Image Search and YouTube hunting for obscure information and encountering unexpected results is a very common one. My blog serves as a visual representation of this phenomenon. This ability to endlessly drift from one topic to the next is the inherently fascinating quality that makes the internet so amazing. A NOTE ON THE PROCESS Just wanted to add a note on how I make this blog, as I have seen people wonder the same couple things frequently. Firstly, lots of people ask if it's a algorithm or something. It's not! Just me searchin google. Secondly, a lot of people assume this blog is therefore made predominantly by using the "visually similar" function on Google Image Search, which is a totally reasonable thing to assume. While I definitely use that function in my attempts to search thoroughly (and love it for it's own beautiful results), surprisingly little of the piece is actually constructed using it. Visually Similar appears to employ an algorithm based mainly on color percentages in an image, and as I'm Google is based more often in conceptual similarities than color-wash similarities, my searching is almost entirely relient on keywords rather than searching by image. Of course, there are times when Visually Similar has helped with a transition or section here or there, but overall, it's not the way I work. I hope you enjoyed my first FAQ – Dina Kelberman

anc84 11y ago

Aww man, I thought this really was pure algorithms, not manual curation. It is still a wonderful project but my awe is gone. Would be great to see it as a real bot.

spacemanmatt 11y ago
Write it!

shrimplypibbles 11y ago

My thoughts exactly. What a waste of time

jahnu 11y ago
Urgh... talk about giving tech people a bad reputation as philistines. It's a wonderful, thoughtful project.
- teagoat 11y ago
  It is a wonderful, thoughtful project, but it's an art project. While that's great, I don't come to Hacker News for the latest interesting art project. I think like many people that clicked the link and scrolled through the images, we made the assumption that these were put together using some algorithm that mined Google's 'visually similar' search and then somehow found the best results in order to create a set of images that 'flowed'. It's a little disappointing to learn otherwise, but it's now kicked off a train of thought about how I might automate something similar. :-)
  - dang 11y ago
    There's nothing wrong with interesting art projects on HN. Story diversity is the most important thing to preserve here.
laferrera 11y ago
This comment is a waste of time.
dang 11y ago
We detached this comment from https://news.ycombinator.com/item?id=10338685 https://news.ycombinator.com/item?id=10338685 and marked it off-topic.

arxpoetica 11y ago

Somebody needs to automate this.

martin-adams 11y ago
Like with Mechanical Turk right?
- rcpt 11y ago
  Could end up like https://vimeo.com/19760997 https://vimeo.com/19760997

abrichr 11y ago

I predict that someone will make an algorithmic version of this and post it to HN within a month.

hantusk 11y ago
There is already an algorithmic version like this. Video here: https://vimeo.com/34949864 https://vimeo.com/34949864
- abrichr 11y ago
  Interesting. It seems to alternate between being "stuck" in certain conceptual valleys, and jumping between different concepts quickly. Perhaps an effect more similar to the op could be achieved by enforcing some sort of smoothness and similarity constraints.
dang 11y ago
Please do!

jstanley 11y ago

How to automate this: - pick a seed image and then: - run one of those "caption generating" algorithms on it (are any open source? what are the best ones?) - feed the caption into google images - pick the first result - repeat Probably also wants something simple to prevent cycles and fixpoints.

jpalomaki 11y ago

"Back in the days" image morphing software was somewhat popular. It could be fun project to run these images through some automated tool to create a continuous video of one image morphing to another then to third and so on.

martin-adams 11y ago
You might not need to morph it, just play it back at 24 fps and the results might be quite interesting. Essentially the hard work of interpolating frames has been taken care of by choosing a slow progression of one image to the next.

dfar1 11y ago

The transitions are brilliant!

Dav3xor 11y ago

Some of the break rooms where I work have a computer you can do searches on. There are some Google image searches that just shine like diamonds. Most of them are somehow PBS related for some reason. "Bob Ross" "Fred Rogers" "Ernie and Bert", etc. I started an odd little game of people leaving google image searches on the break room machines. You'd walk in and.. "Hedgehogs!"

kraig911 11y ago

I wonder would happen if they took every n(20) as a start for another similarity pass.

shubhamjain 11y ago

I am oddly reminded of Pollard's rho algorithim[1], you start with a base image and only choose the first image that comes up in the result, in the end we will eventually return to an image that we have already traversed. I wish someone could find the base images that result in the shortest and longest cycles. [1]: https://upload.wikimedia.org/wikipedia/commons/4/47/Pollard_rho_cycle.jpg https://upload.wikimedia.org/wikipedia/commons/4/47/Pollard_...

accounthere 11y ago
My guess is the shortest would usually be the very next one since Google Image is good at finding the same image file in different websites. There would have to be a constrain for how close to the last image the next one is or you'd end up in a one image loop.

digitus 11y ago

We're working on something really similar to help people explore visual content. https://posters.dreamitget.it/ https://posters.dreamitget.it/ https://watches.dreamitget.it/ https://watches.dreamitget.it/

MrBra 11y ago

I'm assuming porn is filtered out. Given that porn represents the vast majority of Internet content, it would be curious to see how quickly a sexually explicit image would pop up, and how funny the connection with the previous non sexual image would be.

MrBra 11y ago
Oh, it's not an algorithm. Give me back my upvote :)
eonw 11y ago
i don't think porn constitutes the vast majority of internet content by any means, or even the majority at all.
- MrBra 11y ago
  It seems that according to modern estimates this thing about porn being the greatest part of the Internet is no way close to reality and was more of just a legend that got told over and over and ended up being accepted for truth. But honestly I myself don't know and can't be bothered to investigate more. I just know there's enough of it :)

gloves 11y ago

This is cool! I love how on Google when I do an image search on myself I can go from Brad Pitt one minute, to an old man the next. That's not a complaint though. The technology is always improving and I live in constant wonder of the things people can create from nothing :)

Cerium 11y ago

Very cool project. As an interesting side effect I'm sure Dina learned a lot of obscure trade specific words while working on this.

luluganeta 11y ago

This work is from the same year, also really interesting: http://sebastianschmieg.com/searchbyimage/ http://sebastianschmieg.com/searchbyimage/

accounthere 11y ago

It would be nice if you could change the seed image. Someone needs to automate this.

danschumann 11y ago

Kevin Bacon

danschumann 11y ago

Kevin Bacon

hellbanner 11y ago

Reminds me of http://translationparty.com/ http://translationparty.com/ which converts between Japanese & English using google translations

hellbanner 11y ago

A few scrolls down there is "This video was removed by Youtube for privacy reasons".. does the related image algorithm still use what image was there, for calculating next image?

justinator 11y ago
There is no algorithm, this was done by hand.

davidhariri 11y ago

This is so cool!

socialmediageek 11y ago

This is cool. It could become a great business if there are links to buy those items. Like Amazon's similar items (which is still within Amazon) - this may help several businesses.

yazaddaruvala 11y ago

When I read the title "I'm Google - ..." a very small part of me was really, really excited that maybe, just maybe, some system within Google became self aware, learnt english, and wrote a blog declaring its identity.

tajen 11y ago
Man, you've made my day. I've always imagined the first PR about transient AI as "Dr. has applied for a patent for Y" or "How can Siri answer this, although she only knows that". But the idea of a system becoming transient and Google employees not knowing where it is stored inside their datacenters would be plausible. Assuming AI is possible. Hopefully the first thing this AI will declare is whether it stores data on its EU server.

cognivore 11y ago

I love the one that goes from cookie dough to car plowing through the sand dune!