Comments - TinEye Image Search Technology - TechHui2024-03-28T19:28:28Zhttp://www.techhui.com/profiles/comment/feed?attachedTo=1702911%3ABlogPost%3A16266&xn_auth=noHey Brooke - I really like Vi…tag:www.techhui.com,2008-07-03:1702911:Comment:166652008-07-03T04:07:39.535ZDaniel Leuckhttp://www.techhui.com/profile/dleuck
Hey Brooke - I really like View Search Hokkaido. It sounds like a fun project!
Hey Brooke - I really like View Search Hokkaido. It sounds like a fun project! There are government-backed p…tag:www.techhui.com,2008-06-30:1702911:Comment:163192008-06-30T04:53:52.240ZBrooke Fujitahttp://www.techhui.com/profile/BrookeFujita
There are government-backed projects in France and Japan pursuing next-generation image searching. Here's the one in Japan: <a href="http://www.view-hokkaido.jp/">View Search Hokkaido</a> (I developed a web-service driving the back-end).<br />
<br />
@Dan<br />
> Given the enormous computing power brought to bear by Google to index HTML,<br />
> I wonder how it would be possible to build a content based image search that covered<br />
> a substantial portion of the web in a cost effective manner.<br />
Good insight,…
There are government-backed projects in France and Japan pursuing next-generation image searching. Here's the one in Japan: <a href="http://www.view-hokkaido.jp/">View Search Hokkaido</a> (I developed a web-service driving the back-end).<br />
<br />
@Dan<br />
> Given the enormous computing power brought to bear by Google to index HTML,<br />
> I wonder how it would be possible to build a content based image search that covered<br />
> a substantial portion of the web in a cost effective manner.<br />
Good insight, especially the bit about Google.<br />
<br />
One of the driving forces behind this View Search Hokkaido and its wider-ranging parent project <a href="http://www2.igvpj.jp/#top">Information Grand Voyage</a>, is to take on the search giant (singular or plural??) vis a vis image searching. Images can be machine-processed and categorized by n-dimensions (colors, tones, shapes, edges, amount of noise, etc.), and then like images can be grouped by calculating euclidean distance. Very processor intensive, yes, but if these processing tasks can be broken down in just the right way, they can be attacked in parallel by using such libraries as Yahoo!'s <a href="http://hadoop.apache.org/core/">Hadoop</a>, or maybe Google's <a href="http://labs.google.com/papers/mapreduce.html">MapReduce</a>. Or you could simply build your own pc cluster (virtual, even) to do the grunt-work.<br />
<br />
At least, that's the theory as I grokked it.<br />
<br />
My $0.02 Hey John & Scott,
I got…tag:www.techhui.com,2008-06-29:1702911:Comment:163122008-06-29T21:43:23.672ZDaniel Leuckhttp://www.techhui.com/profile/dleuck
Hey John & Scott,<br />
<br />
I got an account on TinEye yesterday and gave it a spin. Per the article Scott referenced, it works well with photos and complex illustrations even when they have minor modifications such as embedded text or certain regions changed. I wonder how well it does with photos that have had subtle filters applied to the entire image.<br />
<br />
<i>John: The main challenge we found were (1) resource utilization and (2) flexibility. Not surprisingly these system can consume very serious…</i>
Hey John & Scott,<br />
<br />
I got an account on TinEye yesterday and gave it a spin. Per the article Scott referenced, it works well with photos and complex illustrations even when they have minor modifications such as embedded text or certain regions changed. I wonder how well it does with photos that have had subtle filters applied to the entire image.<br />
<br />
<i>John: The main challenge we found were (1) resource utilization and (2) flexibility. Not surprisingly these system can consume very serious amounts of computing resources, making the business case for offering such services challenging and often unrealistic.</i><br />
<br />
That makes sense. Given the enormous computing power brought to bear by Google to index HTML, I wonder how it would be possible to build a content based image search that covered a substantial portion of the web in a cost effective manner. Daniel and John, that is real…tag:www.techhui.com,2008-06-29:1702911:Comment:163092008-06-29T20:07:55.697ZScott Murphyhttp://www.techhui.com/profile/ScottMurphy
Daniel and John, that is really interesting and thanks for sharing. Maybe the limitations won't allow the technology to be wide spread but I hope to see it grow and be applied to different websites.<br />
<br />
I later found <a href="http://labs.ideeinc.com/visual/#id=292834;">this</a> by the same company. I thought it might be a cool idea if someone applied this technology to a dating site where you can narrow users based on similar "looks". It might be applicable for a stock photography site as well.
Daniel and John, that is really interesting and thanks for sharing. Maybe the limitations won't allow the technology to be wide spread but I hope to see it grow and be applied to different websites.<br />
<br />
I later found <a href="http://labs.ideeinc.com/visual/#id=292834;">this</a> by the same company. I thought it might be a cool idea if someone applied this technology to a dating site where you can narrow users based on similar "looks". It might be applicable for a stock photography site as well. I did research and testing in…tag:www.techhui.com,2008-06-28:1702911:Comment:162752008-06-28T21:57:17.866ZJohnhttp://www.techhui.com/profile/John
I did research and testing in this space not too long ago. TinEye is an example of <a href="http://en.wikipedia.org/wiki/CBIR">Content Based Image Retrieval</a> and there are a number of companies working on such solutions.<br />
<br />
The main challenge we found were (1) resource utilization and (2) flexibility. Not surprisingly these system can consume very serious amounts of computing resources, making the business case for offering such services challenging and often unrealistic.<br />
<br />
The flexibility…
I did research and testing in this space not too long ago. TinEye is an example of <a href="http://en.wikipedia.org/wiki/CBIR">Content Based Image Retrieval</a> and there are a number of companies working on such solutions.<br />
<br />
The main challenge we found were (1) resource utilization and (2) flexibility. Not surprisingly these system can consume very serious amounts of computing resources, making the business case for offering such services challenging and often unrealistic.<br />
<br />
The flexibility aspect is an issue of how broad a range of image types the system can handle. For instance, Dan mentioned issues with line art. These systems often have issues with different shapes, patterns, color combinations etc. Obviously it depends on how the system is optimized but there are some serious trade-offs in making a general purpose image searching solution.<br />
<br />
One UI technique I have seen use is graphic tools that lets the user select a portion of an image for matching. For instance, you might have a picture with a person in yellow raincoat standing next to the golden gate bridge. The UI allows you to mark a rectangle around the person in the raincoat. There are other UI techniques as well but the general point is to allow human intelligence to relive the strain on the underlying matching system.<br />
<br />
It's a great concept but will likely work best in narrow, specific applications like product searching such as <a href="http://like.com">like.com</a> offers. Great find! TinEye looks very…tag:www.techhui.com,2008-06-28:1702911:Comment:162692008-06-28T19:17:53.481ZDaniel Leuckhttp://www.techhui.com/profile/dleuck
Great find! TinEye looks very interesting. I wonder how they calculate image distances for partial matches and how they index. Akimoto-san's comment that it doesn't work well with logo type line art is interesting. I assume its optimized for photographs and complex illustrations.
Great find! TinEye looks very interesting. I wonder how they calculate image distances for partial matches and how they index. Akimoto-san's comment that it doesn't work well with logo type line art is interesting. I assume its optimized for photographs and complex illustrations.