Image Search via D'n'D

[plasticman]

To my surprise the interwebs just got a new cool feature: Google image search allows using your own images as an input instead of words.

 

It works via drag and drop, so easily that I stumbled into it by accident. Here is an example with a duck I shot myself and never put on the internet anywhere.

 

 

It's the only test I did so far. Will be interesting to see how it works on pictures that are already out there and if they implemented some sort of privacy protection.

 

That's actually why I am surprised about it, I faintly remember reading some Google statement, that they wouldn't let you do this, upload a picture (portrait of your neighbour, friend, foe) and let reverse image search identify that person for you.

 

When it is possible, I am even unsure if I should like it or hate it. It is propably the lesser evil to give this kind of tech to everyone, not just to intelligence services, the police, etc. -- facebook of course...

Edited August 1, 2011 by plasticman

[jaxa]

These guys have been doing it for some time now: http://www.tineye.com/

 

Key differences will probably be improvements to the similar (as opposed to "exact") image search and a manifold increase in speed with Google's 900,000+ servers. As well as the fact that Google won't be going under anytime soon.

 

Your reasoning kind of reminds me of the ongoing Street View controversies. In most jurisdictions it's legal to stand in the street and photograph someone's home. However if you do that 150 million times and upload the results, suddenly you have what is considered a game changer in terms of privacy, and not every jurisdiction has rolled with it. What's clear though is that not all expectations of privacy are warranted. Whether it's street view or maybe some encrypted archive/data you leaked or whatever, or your image as you are recorded passing by 100s of cameras a day, your privacy will most likely be diminished via automation.

 

As for Google's people restriction, I'll have to check this out, because I'm not sure it is feasible. I mean, certainly it is possible to have an algorithm that tells you, "this image contains human faces." Just look at Facebook auto recognition of where faces are in an image (they draw a rectangle around it and encourage you to tag it). The real problem is that Google can't easily decide if the face you uploaded is a celebrity or a regular Joe, without collecting more data and making some mistakes. So I actually doubt there is any kind of "don't upload your neighbor" restriction. Also, does the granularity of the similar image search really pick up different angles of the same person (with different lighting and surroundings, etc.)? Back when I tested Google's SIS it worked well with say, photos of cars or things with white backgrounds. I don't remember how it worked with people and whether the people were the same. That said, Google might have some privacy measures in mind. If you remember, Facebook tried to introduce an algorithm that both recognized where faces are in an image and then autotags each one with the actual name of the person. This caused a firestorm of protest. Around the same time, Google announced that they had developed the same technology independently but had sealed it away forever!

 

Lastly I think you'll find that there is a lot of lesser evil to go around, although it is unavoidable. For example, the NSA and FBI probably already use similar tech, and a good rule of thumb is that the NSA is always 15 years ahead of the public in terms of encryption (probably extends to other mathematical algorithms like facial recognition). But I doubt that most local police have these types of capabilities. So they are getting them along with the general public, in a fairly user-friendly way, and they are more suited to exploiting these capabilities. Although even local police may have more tools than you can possibly imagine. It can be a double-edged sword however.

Edited August 1, 2011 by jaxa

[jaxa]

Part of my underlying argument is that a determined individual can commit some of the same "mass privacy violations" that a faceless corporation can. Remember the Google Streetview WiFi sniffing/logging controversy where Google was excoriated for storing publicly broadcasted information? Well somebody built a drone that can crack WiFi networks, etc. all while flying around. Oh yeah, the updated story is that this year these guys can intercept cell phone calls by having the drone pose as the tower. Presumably they made the thing fly better too.

 

http://blogs.forbes....on-cell-phones/

Edited August 1, 2011 by jaxa

[plasticman]

If you remember, Facebook tried to introduce an algorithm that both recognized where faces are in an image and then autotags each one with the actual name of the person. This caused a firestorm of protest. Around the same time, Google announced that they had developed the same technology independently but had sealed it away forever!

Not like forever: face detection went into google's picasa and you can use it to (auto-)tag your family pictures on your local hard drive.

 

The question is: will we be able to run it not only locally but on google's own omniscient data base (them internet himself)?

 

I thought that was the case now, but my very few tests seem to indicate it's not. Or not yet. Google's moral standards may become a little blurry when there's competition.

[jaxa]

As long as the T&C is subject to change, so are the moral standards.

[jdude]

I don't understand the point of searching an image. If you have the image, why search it?

[jaxa]

1. Find similar images, such as the same subject at different angles.

2. Find loosely similar images, whenever it is useful (call it cold searching for something you may want to use)

3. Find the exact same image, at a higher resolution.

4. The above tasks can only be approximated with text search terms. For example, if an image on the Internet has no or no useful text associated with it, it will not be considered relevant for the purposes of a textual image search. However, by searching for similar images, any indexed image can be algorithmically compared to the input image. Similarity is only the most obvious means of comparison, perhaps there are other ways (for example, search by color range to find similarly vibrant images, etc etc.)

[plasticman]

The idea is to provide a different input method for search in general. It tries to provide useful links in the first place while offering visually similar images as an option.

 

Seems to be quite handy when you need a higher resolution of an image you already have, or -- and this is powerful -- you like to identify images. Art, for example.

 

If your input is RIPPOFF4.GIF as provided in TMA, it discovers the original source for you.

 

I haven't seen a feature like this before. Next step will be you hum a melody and google tells you which song it is.

 

[Aprilsister]

Just used this and it worked nicely.

 

Next step will be you hum a melody and google tells you which song it is.

 

This I've seen and used and it was years ago to boot, it wasn't google (might even have been so long ago it was pre-google!). It did work. I've long since forgotten the site. ...Google it