Milabra, a new image recognition company offering other websites a B2B solution, has launched to the public. In conjunction with the launch, the company has also announced a $1.4 million funding round with a number of private investors contributing. Milabra’s engine is capable of identifying a broad range of images, “from puppies to porn” (their words), and can be applied in a variety of ways like adult content prevention and automatically tagging images with searchable metadata.
The company’s image recognition engine is supposed to mimic some natural tendencies seen in the human brain (Milabra’s team includes PhD neuroscientists from Columbia University). Rather than trying to identify photos based on their similarity to a template image (for example, using a portrait of the Mona Lisa to find identical matches or slightly modified versions of the famous painting), Milabra can “learn” to recognize certain scenes or objects without having to refer to a source image. So when searching through a library of images for dogs, Milabra doesn’t need to constantly compare each image with its database of known ‘dog’ images – instead, it can look for traits that it has learned to associate with “doggyness”, leading to an engine that is more flexible and speedier.
While Milabra offers a demo version of its software on its website, its primarily focus is on creating solutions for other businesses (you can view a list of partners the company has already signed here). Clients are able to manage their photos through comprehensive CMS backend, which can be customized on a per-customer basis.
The company reports an average of 20ms needed to identify each image, which it says is substantially lower than most of its competition. It also reports a success rate of 98% when identifying faces, and its pornography filter (which could be used by family-friendly sites to screen user-generated content) is also nearly 99% successful. However, the software has its limits: in order to add a new filter (say, a search that could identify all photos with baseball caps), customers will have to wait a few days for Milabra to implement their requests – there’s apparently no way to teach the application using your own photos.
During the demo I was shown, the company’s engine worked as advertised. A search for beach images successfully identified all photos taken at the beach, regardless of the lightening or camera orientation. Of course, this was using the company’s own database so it’s tough to tell how accurate the engine is in real-world scenarios. Milabra is by no means the first startup to tackle image recognition – there have been stream of similar solutions that have tried and failed, leading some companies to give up on automation and crowd source the identification to bored humans. So while Milabra may seem to work well, the real test will lie in how many major partners the company is able to sign (Milabra says some major deals are on the way, though they wouldn’t go into details).
Pretty cool.
I guess there is a reason Like / Riya engine never caught up.
The ‘learn’ feature is definitely essential for any upcoming search engine technology.
be nice if the site linked. im getting an error link on firefx and explr.
Unique but boring really!
cooljobsonline
http://tinyurl.com/7uj5ay
Beats your lousy site anyday.
I take your Yasser and raise you an Ismail.
’Puppies To Porn’
that is what a layperson says thinking they are clever. Substitute sport for porn and realise what an idiot statement it is.
call me back when it can detect double ended figging amateur bukkake dogging and I have a multi-billion dollar market for it
(look up figging and tell me how it can detect that)
the generica of web2.0 logos continues.
Apparently if the name ends in bra, you have to use the round-end font… like zimbra
http://www.zimbra.com/
Call me skeptical but I don’t buy any of their claims. I doubt their algo would work well in 1M image sample size. There would be a ton of false positives and false negatives.
They have a carefully crafted data set in their demo.
Hmm, I cant find their demo service?
Also, can they find images on other traits (e.g. location, type) using filenames and tags?)
Say I want to find travel related pictures of Miami that have the beach in them. Do they have the ability to do this? Also can I just specify tags that I want (e.g. beach, landmark, scenic shot) or do I have to upload an image for them to find similar ones to?
Did anyone else catch that they are using… MATLAB?! (from the demo video)
Ahh, I really don’t know what to say to that… do any of you?
It’ll be an outstanding product.. if it works as advertised in a live environment..
That technology would potentially be able to rake in billions of dollars of profit for its use..
they say in their website that they are the winners of onmedia 2009 and when you click on it you see that they were only contestant…
LOL.
Congrats on the big onmedia win!
No, we *were*a contestant, and we won the Top 100. AlwaysOn put out a press release in late Jan. I guess we need to make that clearer on our website.
The proper url to the OnMedia results is this: http://alwayson...link/post/30893
And they were on the list …
Demo Service was online last week but is no offline
… wonder why ???
As far as the linked articles states they only were nominated. Quantcast has won in their category.
I suggest the company investigate their IP position before investing too much time or energy in this platform. Just a friendly recommendation…
I am the chief scientist at Milabra. Let me clear up some of the FUD going on here:
- Of course the technology works, or we wouldn’t have started the company. In general, if you think that this cannot be done, then you are completely clueless: object classifiers have been made for more than 10 years now at leading CS labs around the world. Our secret sauce is not that we can make them, but that we can deploy them quickly, and make them work FAST (30-40 times faster than our competition).
- In the response to the matlab question above, yes, the demo on the website uses matlab. That is where we develop classifiers, before we deploy them on (much) faster C based systems. But if you think matlab is somehow inferior, then you probably haven’t used it since school. We could deploy everything in matlab and still crush the competition, both in terms of speed and accuracy. Matlab has come a long way in the last 10 years!
- The interesting point raised above is by the Y C Wee, who wondered if it will work as advertised in a live environment — that is indeed the true test of our mettle!
- There were 800 companies reviewed for the AlwaysOn Top 100 list, 350 shortlisted, and 100 selected. We were a winner in the “Technology Enablers” category. Do your own research.
Onward and upward!
Thanks for your response. However, there are some things I don’t understand. You first established that C is faster and more practical for your system, over MatLab. Then, you questioned my claims of MatLab being inferior? What?
I actually have used MatLab quite a bit lately in the last few months. Performing functionally-identical operations, such as something simple like cellular automata, with Matlab and C side-by-side, had totally different processing times. Obviously Matlab has its uses, like crude prototyping. All I did was question the use of it for something that’s already getting this sort of media coverage. Obviously you had to have had something else, other than your basic MatLab implementation, in order to come up with the 20ms figure from the article? The returned images from the query in your demo came up at a speed of about one per second—which I wouldn’t qualify as FAST, or compare to the 20ms figure mentioned in the article.
Don’t take TechCrunch comments personally. For me, reading the audacious claims of it being lightning-fast, and then seeing the demo video using MatLab, just seemed overly contradicting.
(ps. Also, I’d venture to say that no one here really gets how a magical numerical “threshold” defines “beach” or “dog”.)
Naveen,
Thanks for your comments above. But no mention of your IP position? Curious…
There is another company that offers a product that sounds similar to Milabra called SoarSpace.
It seems that they tag videos as well as images. The demos that I saw on their web site look very interesting, as it seemed to match video content to relevant ads based on the content of each scene within the videos. Has anyone else heard or dealt with them?
Please see http://www.picscout.com.
They have a better solution
The claims seem to be realistic apart from the speed they mention.
@Naveen – when you said 40 times faster – i hope you meant all the companies you are referring – right !
Classifiers like beach are quiet easy to write and implement – lets talk about something like a crowded market or a park [not jungle] – I hope you can add your $0.20 on the same.
We too are a company in the same domain and have been recently incubated by IIM Ahmedabad business incubator. [ Visit http;//www. picporta.com ]
We to have classifiers from beach to automobiles and currently working.We too offer search time less than 10 msec.
Keep me posted on your findings
-
Pulkit Gaur
Founder and CTO
Picporta Pvt. Ltd.
Looks reasonably good (though not amazing) considering the current state of the art!
I guess the main reasons for most of the above negative comments is because the video is “not” spoken in an American accent and the photos are also not all American…
Add another video with these corrections and I bet there will be tons of positive comments!!