Picollator Image Recognition Search
There are a bunch of companies trying to figure out how to get a machine to be correctly identify images just by looking at them. We've profiled a bunch on ReadWriteWeb before: Eyealike, Mugr, Riya/Like, and Pixsta. Generally the tech demos for these visual search technologies fall into two categories: shopping search or facial recognition. The latter is on display in a new Russian startup we came across this week called Picollator.
We noted last April the problems with current gen image search, which relies on keyword analysis to determine what a picture's subject matter is. While Google and the other major search engines have gotten pretty good at using clues in surrounding text to identify what's in a picture, they're not perfect and often serve up lousy results -- especially when there is a small sample size.
So a search technology that actually looked at a picture and could understand what was in it -- or at least match like images -- would have a large number of practical applications. Picollator's online demo is based on software of the same name from parent company Recogmission. The demo allows users to upload an image of a face to the site, and then attempts to match it to similar images it has gathered from around the web.
One of the main issues we've noticed with most of these facial recognition products is their inability to deal with my facial hear -- sometimes even thinking that my dark beard is really an indication of my skin tone. So I was excited when Picollator matched me to a picture of Russell Crowe. I don't really look like Crowe, but believe it or not, people actually used to tell me I did back in college.
Unfortunately for Picollator, subsequent tests with different images were less successful. It had trouble dealing with an older picture of mine (though to be fair it had less face and more body), and also didn't respond well to a black and white picture. And if their software really is accurate, I am dating a girl who basically looks like any celebrity that smiles.
What that means is basically that image recognition software still has a long way to go. Certainly, these technologies will improve over time, but for now, the flawed approach of text based image search from Google, Yahoo! and the other big search engines is probably the best way to go.
Share
Comments
Subscribe to comments for this post OR Subscribe to comments for all ReadWriteWeb posts
I think that the problem with image recognition is that in the web technology context, research proceeds isolated from research in other fields. The obvious one is in medical imaging, where using measures of entropy to compare two images has long been a standard technique for preparing two images for mutual registration (and deformation), so that (for example) an image acquired with an MRI can be directly overlain onto one acquired with a CT. Here's just one of hundreds of review papers on the topic: http://tinyurl.com/6on6en
In principle, image recignition would simply be a systematic approach. Start with the source image, and then do an entropy comparison against a stock collection of reference images (perhaps on Flickr, accessed via API). Inspect the tags on the ten highest matches to determine the categories, and then rerun the comparison against 1000 more images in each of those categories. Repeat until the improvement in entropy falls below some threshold. The advantage of such an approach would be that it would inherently tap into two sources of information: genuine algorithmic computation, and the folksonomic meta-information supplied by human beings (in a distributed and efficient way). Of course the processing power required would be immense, but in this age of cloud computing and massive server farms, is that really anything more than a minor logistical detail?
Swedish startup Polar Rose is another example.
@Aziz,
I think the problem with what you are saying lies in finding a suitable "entropy comparison". For 3D objects imaged under all sorts of changes like illumination, shadows, occlusions and so on; it is a very challenging problem to find a good comparison measure. Indeed much of the last two decades of Face Recognition has been spent on doing precisely this.
Obviously the problem of Face Recognition has enormous commercial potential, but the sad truth is that the results in theory never match the practice because of the controlled research conditions. I think the huge commercial interest would drive people to find a good solution (not perfect) within the next couple of years.
Josh,
thank you for reviewing Picollator project in your post.
We in Recogmission fully understand that Picollator has a long way to go. Picollator Beta - is a step one on our way, so you will soon hear about further steps in image recognition field.
We are now developing a new approach to search for information in the Internet based on visual query instead of textual query. Soon we will provide information about step 2.
i'm actually looking for a service that does somewhat the opposite to the above: match content to a given photo, i.e. i upload a picture and get results where the picture is taken from, e.g. from which book, which internet site, which of my friend's flickr page, etc.
does anyone know of such service?
balazs