Software Combines Thousands of Online Images Into One That Represents Them All

Zothecula writes If you're trying to find out what the common features of tabby cats are, a Google image search will likely yield more results than you'd ever have the time or inclination to look over. New software created at the University of California, Berkeley, however, is designed to make such quests considerably easier. Known as AverageExplorer, it searches out thousands of images of a given subject, then amalgamates them into one composite "average" image.

Comments

First post

[KamikazeSquid]

Can we use it to create an amalgamation of the "average" first post on a /. article?

My fear

[Anonymous Coward]

I dread that the average of all internet photos will be a pornographic picture of a woman with a penis.

Scientific justification

[sinij]

So this is how a bunch of scientists justified browsing for pron and cat videos all day long? Yes, we heard this before - they are working on average algorithm for images.

You'll be impressed

[i kan reed]

With this complex algorithm that takes a fuck-ton of image data and produces for you: something that is almost impossible to tell apart from applying the blur filter on the original image.

Yay! Average!

[Tree131]

Suddently there's a surge in searches for "Average Penis"

Actual article

[Anonymous Coward]

Here's a link to the actual article, rather than the useless link provided:

http://newscenter.berkeley.edu/2014/08/14/average-image-for-big-visual-data/

The video was pretty interesting!

The Average Cat

[sfsp]

So...what the software demonstrates is that if you line up all the pictures of cats by centering them on their noses, you will CLEARLY see...

...that the average cat has a nose.

The rest is blurry and remarkably uninformative.

There needs to be a LOT more intelligence, either machine or human, applied to this before it is remarkable.

Re:The Average Cat

[The Raven]

You may have read the article (dubitable), but you didn't watch the video or read the SIGGRAPH paper. They demonstrate a browsing tool that enables you to, for example, find an average nose nearly instantly. You can then filter the thousands or millions of images to find specific cat breeds, poses, situations, or colors in seconds.

The tool is called average explorer, and it allows a user to interactively explore a vast set of image data quickly and efficiently. The one picture you describe was a single click in the explorer.

You did the equivalent of saying "Wow, I can make a black dot on a white canvas. That's not very exciting." when presented a single click with a single tool in Photoshop.

Automatically means no control

[techno-vampire]

If this software searches out all images of a subject and averages them automatically, that means that there's no human control over which images to use and which to reject. Imagine what would happen if you were to let this program loose to create an average image of Shirley Temple. She started in films at the age of three and reached the age of 85, and the software would create an "average image" by mixing images of her as a small child with ones of her as an elderly woman. Even worse, there's a non-alcoholic cocktail named after her, and pictures of it would almost certainly get included.

Re:Frist psot

[Zero__Kelvin]

Yes. That looks about correct.

it works!

[nblender]

I typed in "douchebag" and it showed me a picture of some guy driving an Audi A4.

I tried it

[Solandri]

Here's what I got when I gave it every pic in my photo library.

But seriously, I've seen the same technique used to discredit a movie of a UFO shot on 8 mm film. If you just watch the movie, you see an elliptical blob flying. Someone scanned the blob from each frame, aligned them, and averaged them. The increased contrast (bit-depth and resolution basically) let you see that the elliptical blob was more a diagonal prism, and that there were dark features underneath it. Basically it was a Cessna with the sun reflecting off the top of the wing.

Three Dee

[Tablizer]

I'd like to see one that constructs a 3D model. Perhaps it could use a genetic algorithm (GA) to breed a 3D model that can best represent the most actual specimens of the target object type.

It may be a lot of computations, however, because one is not just running genetic algorithms, but also rotating all the candidate 3D models and lighting conditions to see which best fits the actual specimen images PER GA candidate PER specimen. Perhaps a 3D thumbnail version can be used to for initial placement estimations to be fine-tuned with a fuller model.

Then you got spot and texture variations within specimens. You have to model varying textures, not average them out. But even if it ignores texture & spots to simplify things, a 3D shape model result would be cool.