Author Topic: Finding images with similar content that also look similar (cmts please) (Read 9899 times)

pixo · « **on:** October 18, 2013, 03:40 »

Finding images by keywords is difficult. Keywords often are incomplete or overloaded and might suffer from translation errors. Visual image search finds similar looking images without understanding the meaning. Therefore it will often find similar looking images showing different things.
To deal with this problem we have developed a new approach combining visual and keyword search:
http://fusion.pixolution.de/

How to use it: Start your search by typing a keyword at the top of the page. Then move your mouse over an image you like to start a similarity search. You can focus more on keywords (the meaning) or on visual similarity. Here are three possible results searching for dandelion images:

I'd like to hear your opinion on this new way to search for images.

P.S.: I discovered the Microstock Group forum only recently and as a newbie I have posted a similar topic in the Newbie Discussion, which does not seem to read by too many people. So sorry for crossposting.

ShadySue · « **Reply #1 on:** October 18, 2013, 04:51 »

Quote from: pixo on October 18, 2013, 03:40

P.S.: I discovered the Microstock Group forum only recently and as a newbie I have posted a similar topic in the Newbie Discussion, which does not seem to read by too many people. So sorry for crossposting.

You may not have realised that this forum is used mostly by people who are selling content, though some also buy content. So this may not be the best marketplace for your product.

pixo · « **Reply #2 on:** October 18, 2013, 04:59 »

Quote from: ShadySue on October 18, 2013, 04:51

You may not have realised that this forum is used mostly by people who are selling content, though some also buy content. So this may not be the best marketplace for your product.

You are definitely right, but we need to know how people do search for images. And therefore feedback to new image search approaches is quite important for us.

Have you ever tried to find your own images among millions of other images without using the photographer's name?

ShadySue · « **Reply #3 on:** October 18, 2013, 05:05 »

Quote from: pixo on October 18, 2013, 04:59

Quote from: ShadySue on October 18, 2013, 04:51
You may not have realised that this forum is used mostly by people who are selling content, though some also buy content. So this may not be the best marketplace for your product.

You are definitely right, but we need to know how people do search for images. And therefore feedback to new image search approaches is quite important for us.

Have you ever tried to find your own images among millions of other images without using the photographer's name?

It's easy for my own images as I know my own keywords, but I can see it might be a problem if I was specifically looking for someone else's. [1]OTOH, once I had found them once, I could bookmark their port.
[1] In reality, it isn't difficult, because if I know a specific person's body of work, I can search via a relatively unusual keyword or keyword combo and find a few images and get their port that way.

Kerioak~Christine · « **Reply #4 on:** October 18, 2013, 06:21 »

I have just looked at your demo version - I am curious as to where are the images coming from as there are a lot of mine there, many of which have not been available online for some time, nor were they all licenced.

Sean Locke Photography · « **Reply #5 on:** October 18, 2013, 06:33 »

Quote from: Kerioak~Christine on October 18, 2013, 06:21

I have just looked at your demo version - I am curious as to where are the images coming from as there are a lot of mine there, many of which have not been available online for some time, nor were they all licenced.

They said in the last post they are from Fotolia.

pixo · « **Reply #6 on:** October 18, 2013, 06:52 »

Quote from: Kerioak~Christine on October 18, 2013, 06:21

I have just looked at your demo version - I am curious as to where are the images coming from as there are a lot of mine there, many of which have not been available online for some time, nor were they all licenced.

We are using 22 million thumbnails from fotolia to demonstrate the search possibilities. If you move your mouse over an image then you can click "Show Details" to see a larger version. There is also a link to the fotolia site with further information about the image.

Kerioak~Christine · « **Reply #7 on:** October 18, 2013, 07:19 »

Most of mine have been removed from Fotolia - some quite a few months ago which is concerning that the smaller images are still in use. It is so easy to get your images listed, much more difficult to get them removed from everywhere they have allowed them to be used.

They now only appear on my site below.

The software work nicely though

ajt · « **Reply #8 on:** October 18, 2013, 08:51 »

Very nice concept, I like this 0-100% bar.
I am developing something similar at http://symbiostock.info, but it will be a bit different solution

pixo · « **Reply #9 on:** October 19, 2013, 06:54 »

Quote from: ShadySue on October 18, 2013, 04:51

You may not have realised that this forum is used mostly by people who are selling content, though some also buy content. So this may not be the best marketplace for your product.

I did comment on this in a previous post.

But just one additional question: Does anybody know what might be the "right market place" for us?
I did not find any group that deals with the problem how image stock systems should be designed or want kind of software the agencies are using.

Thank you.

heywoody · « **Reply #10 on:** October 19, 2013, 07:31 »

From a cursory read of the thread and zero experimentation I'm assuming that the software will generate a sortable "similarity value" - on that basis, the best market would be to market as a component with an interface that could be used by web sites and suppliers of image editing software.

OM · « **Reply #11 on:** October 19, 2013, 11:01 »

Seems to work well (once I realised that the % visual was adjusted by my cursor position. I don't read instructions too well!).

michaeldb · « **Reply #12 on:** October 19, 2013, 21:18 »

The software was uncannily effective at finding similar images. I was impressed.

I wonder how it would work on, say, Amazon. After reading a good novel, let's say historical fiction, I often wish I could find more like it, but sometimes keywords don't work too well. If Amazon let me search for books with covers like the one I just read, I would try it to see what it came up with.

BTW one of your pages says "Visual Arrangement...Inspect large image sets at a time without loosing overview."
I suspect you mean 'losing' instead of 'loosing'.

pixo · « **Reply #13 on:** October 21, 2013, 08:12 »

Quote from: michaeldb on October 19, 2013, 21:18

The software was uncannily effective at finding similar images. I was impressed.

I wonder how it would work on, say, Amazon. After reading a good novel, let's say historical fiction, I often wish I could find more like it, but sometimes keywords don't work too well. If Amazon let me search for books with covers like the one I just read, I would try it to see what it came up with.

BTW one of your pages says "Visual Arrangement...Inspect large image sets at a time without loosing overview."
I suspect you mean 'losing' instead of 'loosing'.

Thank you for the nice feedback (and the correction

).

In the current demo we only use keyword & visual information to retrieve similar images.
At http://www.picslikethat.com/ (another university project) we have also used the click pathes from previous searches to learn relationships between images. The problem is that you need many users to learn enough relationships. This kind of search approach could also be used for other products.

MicrostockGroup Sponsors

Author Topic: Finding images with similar content that also look similar (cmts please) (Read 9899 times)

pixo

ShadySue

pixo

ShadySue

Kerioak~Christine

Sean Locke Photography

pixo

Kerioak~Christine

ajt

pixo

heywoody

OM

michaeldb

pixo

Related Topics

Sponsors

Microstock Poll Results

Sponsors