pancakes

MicrostockGroup Sponsors


Author Topic: Solution for automatic keywording  (Read 997 times)

0 Members and 1 Guest are viewing this topic.

« on: May 29, 2019, 07:40 »
0
We have launched the new version of Everypixel DAM, a service that will cut your time costs on a microstock distribution routine.

You don't need to spend several minutes to invent the keywords for each file. Try auto-keywording! Neural networks of the Everypixel DAM will select up to 49 relevant keywords for your photos and videos in seconds. Just upload your shoot to the system and click on the "Auto keywording" button to automatically match the keywords to the entire content of the folder. Also, you can to keyword every single document and quickly edit them: you need only one click to delete each keyword.

You can download these keywords in CSV-format or use it for the photo banks via the Everypixel DAM platform. On the free plan you have the storage for 3 GB, 50 submits and 50 requests to auto-keywording.

Try it right now!



« Reply #1 on: May 29, 2019, 14:17 »
+1
The FTP access doesn't work.

"Error -203: miscellaneous error occurred while trying to login to the host"
 
Though I did use the upload interface on the actual website and gave it ago. I uploaded a head shot of a man. The auto keywords were reasonably accurate. I did see a few not applicable keywords. So if one does not care about a few bad keywords then this is a real time saver.

I have wondered on a separate note if it matters that photos are uploaded to stock websites with mostly accurate keywords and a few not applicable keywords. Because it seems as if the stock websites run test on which keywords on different images get a click from real users of the site. So the website learns over time which keywords on different images to either ignore or prioritize.

The photo I uploaded is a man on a green screen not smiling. See attached.

These are the results from the auto keywords
Men, People, Adult, One Person, Males, Caucasian Ethnicity, Portrait, Looking At Camera, Male Beauty, Studio Shot, Copy Space, Young Adult, Looking, One Man Only, Casual Clothing, Confidence, Human Face, Serious, Young Men, Head And Shoulders, Color Image, Only Men, Close-up, Indoors, Adults Only, Real People, Lifestyles, Front View, Smiling, Standing, Facial Expression, Green Color, Colored Background, Green Background, Beautiful, Headshot, Cheerful, T-Shirt, Blue, Masculinity, Positive Emotion, One Young Man Only, Fashion, Beautiful People, Brown Hair, Thinking, Posing, Shirt, Backgrounds

Some inaccurate keywords are - smiling, cheerful, blue, positive emotions, fashion, thinking, posing, backgrounds.

It seems all the inaccurate keywords are towards the bottom of the list, which seems to suggest the list is sorted based upon a confidence ranking. Might be easier to give the user an ability to filter out keywords that have a low confidence score. Google, Microsoft, Amazon's image recognition software all put a confidence number next to the keywords they produce. Such as Men (99), Casual Clothing (70), chair (12). In this example, anything below a 70 could probably be filtered out. 
« Last Edit: May 29, 2019, 14:31 by charged »

« Reply #2 on: May 29, 2019, 14:44 »
0
Did another test image of glacier in Argentina.

Here's the auto keywords:
Snow, Nature, Mountain, Ice, Glacier, Blue, Winter, Cold - Temperature, Scenics, Landscape, Iceberg - Ice Formation, Antarctica, Sea, Sky, Frozen, Outdoors, Water, Mountain Range, Beauty In Nature, Mountain Peak, Cloud - Sky, Polar Climate, Arctic, No People, Lake, South Pole, Extreme Terrain, Travel, Panoramic, White, Ice Floe, Travel Destinations, Tranquil Scene, Wilderness Area, Color Image, Cloudscape, Day, Europe, Melting, Sunlight, Frozen Water, Rock - Object, Summer, Frost, Snowcapped, European Alps, Greenland, Tourism, Idyllic

Inaccurate keywords according to me are...
Antarctica, south pole, panoramic, wilderness area, europe, melting, rock - object, frost, european alps, greenland, idyllic

Otherwise it is fairly accurate. That particular photo already had keywords embedded by me previously. The software should have looked at the keywords within the file instead of ignoring them. Keywords don't just magically appear in photos. If the software had looked at the keywords within, it would have known the location was Patagonia, Argentina.

I'd consider using this service if there was a pay as you go option. Not keen on subscription. The one down side is that this service has no ability to be incredibly specific about location. So if I was going to manually type in Patagonia, I might as well use any free keywording sites online and search for patagonia, select all images from shutterstock that look similar, then pick out the keywords I want. That would be more accurate than this software. Just takes more time. So it seems to be a trade off between being specific and saving time.

Also how about auto title and description. This can probably be done by looking at the keywords and using some logic on how to structure the title.
« Last Edit: May 29, 2019, 15:18 by charged »

« Reply #3 on: May 30, 2019, 04:06 »
0
Thanks for your attention!

The FTP access doesn't work.

"Error -203: miscellaneous error occurred while trying to login to the host"

The FTP access was temporarily suspended due to exceeding the limit of active users. We've raised the limit, so now FTP is able :)

Some inaccurate keywords are - smiling, cheerful, blue, positive emotions, fashion, thinking, posing, backgrounds.

We assume that not all keywords may be quite relevant for your photography. That's why we have implemented a convenient interface for metadata editing and you can delete inaccurate keywords in one click.

It seems all the inaccurate keywords are towards the bottom of the list, which seems to suggest the list is sorted based upon a confidence ranking. Might be easier to give the user an ability to filter out keywords that have a low confidence score. Google, Microsoft, Amazon's image recognition software all put a confidence number next to the keywords they produce. Such as Men (99), Casual Clothing (70), chair (12). In this example, anything below a 70 could probably be filtered out. 

The confidence score may be helpful in some situations, but this can affect that users will be too strict in their approach to filtering keywords. Like in your example, clean up everything rated below 70. As a result, the effectiveness of auto-keywording will decrease, cause a low rating doesn't always mean that the keyword is unnecessary and doesn't correspond to the image.

The software should have looked at the keywords within the file instead of ignoring them.

We're working on it ;)

« Reply #4 on: May 30, 2019, 13:16 »
+3
definitely not ready for prime time - by ignoring existing info like title & desc (which still require separate editing), you miss the most important info about location & subject

so for this image It found:
Outdoors Decoration Architecture Cultures Ornate History Ancient Religion Old-fashioned Metal No People Iron - Metal Nature Spirituality Antique Stone Material Close-up Formal Garden
Wood - Material Green Color Symbol Sculpture Cast Iron
Tree Day Statue Park - Man Made Space Obsolete Single Object
Electric Lamp The Past Architecture And Buildings Asia
Bronze - Alloy Grass Color Image Lantern Backgrounds
Retro Styled Cross Carving - Craft Product Concepts And Ideas
Tranquil Scene Ornamental Garden Summer Europe Steel Macro

my description:
"Bronze Shiva in garden, with blades of grass. Nataraja (Sanskrit: Lord of Dance) Shiva represents apocalypse and creation as he dances away the illusory world of Maya transforming it into power and enlightenment."

so, while it might do ok with simple stock images, on the unique ones, the ones where we need assistance, it fails
 



 

A picture of Taj Mahal w no metadata does tag "Taj Mahal", but also added orthodox,christianity,church,cathedral,russia, turkey, middle east  I know I can delete these but that takes more editing time for each image

overall, while it suggests some keywords i wouldnt have found, most are so generic they wont help much in a search   plus, since it overwrites any existing tags, i'd either have to re-enter, or do yet another edit to add the missing data.  time savings are at best minimal


 
« Last Edit: May 30, 2019, 13:24 by cascoly »

« Reply #5 on: May 31, 2019, 07:22 »
0
so, while it might do ok with simple stock images, on the unique ones, the ones where we need assistance, it fails

We experimented. We measured the sales of photos with auto keywords and keywords selected by a person. The number of sales with automatic keywords was at least 10% higher.

As we found out, this is because people are more likely to look for photos in general terms, rather than by specific names of locations and objects. As a result, the images containing general keywords received more sales than those containing original place names.

« Reply #6 on: May 31, 2019, 07:43 »
+2
so, while it might do ok with simple stock images, on the unique ones, the ones where we need assistance, it fails

We experimented. We measured the sales of photos with auto keywords and keywords selected by a person. The number of sales with automatic keywords was at least 10% higher.

As we found out, this is because people are more likely to look for photos in general terms, rather than by specific names of locations and objects. As a result, the images containing general keywords received more sales than those containing original place names.

Was your experiment done on commercial photos only or commercial VS editorial?

I find editorial content to have more sales when it's very specific.

« Reply #7 on: May 31, 2019, 11:07 »
+2
so, while it might do ok with simple stock images, on the unique ones, the ones where we need assistance, it fails

We experimented. We measured the sales of photos with auto keywords and keywords selected by a person. The number of sales with automatic keywords was at least 10% higher.

As we found out, this is because people are more likely to look for photos in general terms, rather than by specific names of locations and objects. As a result, the images containing general keywords received more sales than those containing original place names.

Was that person someone who didn't know how to keyword? Software can get better with refinement, people get better with experience. I shoot a lot of the same things over and over again because I know in general they sell, but I don't actually know which image will sell because that is sort of up to the algorithm of the stock website to decide which images to serve up on the top of their search results. Your claim of 10% higher sales by auto keywords is in my opinion impossible to prove. And I've already said in an earlier post that I thought that your software was quite good, though not perfect. It lacks the nuisance understanding about the context about some photos that only the original photographer could know. Take my glacier photo, your software did not know it was from Argentina. Realistically it was impossible for your software to know, it was an iPhone photo but the GPS data was missing, I didn't take it out. So I suppose one of the editing software I used took it out. Anyway, I think your claim 10% higher sales is  silly.

Also I assume people search both in general terms and specific terms. A human keyworder who knows the context of where the photo came from would put in both types of keywords. Your software would lose sales from those who are searching with very specific terms.

« Reply #8 on: May 31, 2019, 12:40 »
+1
so, while it might do ok with simple stock images, on the unique ones, the ones where we need assistance, it fails

We experimented. We measured the sales of photos with auto keywords and keywords selected by a person. The number of sales with automatic keywords was at least 10% higher.

As we found out, this is because people are more likely to look for photos in general terms, rather than by specific names of locations and objects. As a result, the images containing general keywords received more sales than those containing original place names.

your conclusion is not the only possibility.  you forgot to control so many variables: who chose the photos to test?  how did you obtain sales info? how big was your sample? how diverse?  did all the images reviewed have sales?  during what time period?

and, 'general' images may sell more, but the competition is greater (by about 11.5% ) so more specific images have a better chance of sales

ShadySue

« Reply #9 on: May 31, 2019, 13:25 »
+3
As we found out, this is because people are more likely to look for photos in general terms, rather than by specific names of locations and objects. As a result, the images containing general keywords received more sales than those containing original place names.
From the agency on which I can see search terms, that's not true. Specific searches are most common, then general, then a small proportion of 'goodness knows what they wanted'. Possibly that's not true of all agencies, I couldn't possibly say.
But anyone with any sense keywords specifically and then outwards to more general terms.


« Reply #10 on: June 05, 2019, 06:07 »
0
Was your experiment done on commercial photos only or commercial VS editorial?

I find editorial content to have more sales when it's very specific.

Our friendly photo production studio tested the algorithm on several commercial shoots.

Keywords for half of the images were selected with Everypixel DAM, for the other half with the help of distribution specialists. They have been working on keywording for many years.

At the end of the month, the sales of the images with auto-keywording were higher.

« Reply #11 on: June 05, 2019, 09:32 »
0
Was your experiment done on commercial photos only or commercial VS editorial?

I find editorial content to have more sales when it's very specific.

Our friendly photo production studio tested the algorithm on several commercial shoots.

Keywords for half of the images were selected with Everypixel DAM, for the other half with the help of distribution specialists. They have been working on keywording for many years.

At the end of the month, the sales of the images with auto-keywording were higher.

Must be not very good human keyworders. There is inherently no reason why a human wouldn't be every bit as good as software, if not a great deal better. Humans understand a lot of context that software for the moment often can't. Even if the software knew some keywords that the human didn't, the human only needs to look once at what the software did, and now the human knows too, and the human still has extra ability to understand context. What is the great promise of software driven keywording is 'speed'. Humans will never do it as fast as software. Multiply that time saving over thousand or tens of thousands of files, that is an enormous amount of time saved. Time is money, thus an enormous amount of money saved. Anyway, I think your software is great, but I also think it is silly for you to say your software keywords are better than human keywords. 

« Reply #12 on: June 05, 2019, 09:52 »
0
Time is money, thus an enormous amount of money saved.

No, good files with ACCURATE, RELEVANT, keywords is money. :)

Getting 10,000 images up with crappy keywords is pointless, which is why we hear so many complaints from people with big portfolios and no sales.

AI can't even come close to an average human brain, BUT, it could be useful as an idea generator - one quick look at the suggested keywords, and then pick out a few that are relevant. This wouldn't save time though, but might be a little bit of help.

I would never ever just go with the suggestion because it will in about 100% of the cases contain errors and miss out on important keywords.
« Last Edit: June 05, 2019, 09:56 by increasingdifficulty »

« Reply #13 on: June 05, 2019, 18:13 »
+1

.

Our friendly photo production studio tested the algorithm on several commercial shoots.

Keywords for half of the images were selected with Everypixel DAM, for the other half with the help of distribution specialists. They have been working on keywording for many years.

At the end of the month, the sales of the images with auto-keywording were higher.

even if it was 1 month after acceptance AND assuming equal quality, acceptance rate and salability , that's not much of a test, esp'ly if sales were only 10% difference --  that could easily be explained by random events


 

Related Topics

  Subject / Started by Replies Last post
4 Replies
2898 Views
Last post February 03, 2009, 12:26
by Anyka
2 Replies
1430 Views
Last post October 16, 2013, 15:14
by pixo
6 Replies
2353 Views
Last post March 30, 2014, 11:51
by mike ledray
2 Replies
851 Views
Last post April 03, 2017, 08:40
by Chichikov
1 Replies
1270 Views
Last post January 16, 2019, 08:27
by Uncle Pete

Sponsors

Microstock Poll Results