We are getting there..slowly. It is not an easy road, but we are getting closer. In the last year or so, we have seen more and more image search companies come out and expose themselves. Even the even mightiest, and certainly the worst, Google Image is thinking about changing its algorithm.
The holy grail is, of course, the end of the keyword based search ( aaaargh !). The first baby step we are currently seeing only focuses on face recognition. For two main reasons :
– A face is always a face, a triangle between two eyes and a mouth, and rather easy for a computer to recognize.
– Between celebrities and relatives, when you deal with image search, the majority of people are looking for either friends, relatives, themselves or more pix of celebrities . There is a huge market.
So here goes many worldwide software engineers claiming image search nirvana. At least when looking for people. I have tested a few, recently, all in Beta and none quite there yet.
TinEye: In a league of it own right now as it does much more than just face recognition image search. Probably the most advanced of all, its limitation is its extreme accuracy. Looking for an image and it will find that image, nothing less, nothing more. Every altered version of it. Great for many, many usages, but a bit limited for those just seeking a similar or inspiration.
Very far away are :
– Polar Rose. Swedish based, it has been full of promises for many years but with disappointing result up to now. Also starting in the face recognition, its algorithm just became clear to me when it recently invited me to Beta test its Plug in. I was expecting a TineEye sort of plug in, but instead, they put me to work. What is up with that ? .
The plug in works like this : every time you are on a web page, it scans every image for a face. It then puts a square around the face and asks you to put the name. As you enter the name, it starts suggesting options. It was always right on the money. At first, I was really impressed, as in ” How does it know ?”. And briefly later on, I recognized the trick. It scan any available text around the image, looks for two words next to each other starting with a capital letter and assumes that should be the name of the person in the pic. All I have to do is confirm. Thus, Polar Rose is currently no more than an elaborate and free version of Amazon’s Mechanical Turk. Putting the community to work. While there nothing wrong, in principal, with this approach, with all these website these days that ask for my free input, I am close to putting full 24 hours days of work, for free.
Their chances of success, using that approach, is minimal. I, for one, will not become a slave to the machine.
– Picollator : Russian by birth, this new image search also looks for faces. You can even upload a sketch and it will try to match it. I tried with this image :
surprisingly got no result. So I went for the TinEye favorite, the Mona Lisa:
.
I got this : 
Some matching results, some very weird ones.
Face recognition and matching can be a very useful tool for the news and celebrity world as photographers and editors could quickly edit film with proper name spelling by scanning the web for matching results. Especially for those B to D celebrities. It is, at this stage, still very sketchy and not quite ready for the big time.
It is also a good indication of how far we still are from a computer recognizing, properly, everything in an image. Right now, we are only scratching the matching part and even that, has it flaws.
It is also very important to know that all these sites are in Beta, meaning far from claiming 100 % accuracy and should not be dismissed until officially proven ineffective. You can try all of them, for yourself :
I am not quite agree with such estimations. Re PolarRose, it is quite correct. The rest is not so obvious. You are right that TinEye limitation is the extreme accuracy (which is not shown in your test, BTW). It is obvious that the system is not the most advanced one because they say they do not understand the image content. They just performs a kind of general analysis = they do not recognize. Re Picollator: did you expected that the recognition system should compare that totally abstract icon with faces? If it does, I would say it is interesting, but I would expect random results from this operation. So it did not. I think it is quite good. I also tested TinEye and Picollator and found that the last could be even more prospective if they collect the bigger database. It searches in more general way trying to recognize indeed. It can rank results – they opened Web search by images in addition to image search by images. Not quite clear results, but the approach can be evaluated as the newest one towards all other search systems. So – let us see.