MicrostockGroup
Microstock Photography Forum - General => Microstock News => Topic started by: dam.everypixel.com on March 22, 2019, 05:49
-
The tariff scale of the Everypixel DAM was completely updated. Now, our neural network for keyword generation is available for everyone! You can use the computer vision to describe photos and videos and then upload it to the stock sites or receive metadata in CSV-format.
— Are you ready to try the AI-keywording? Use the Free plan. You will get a 3 GB file storage, 50 submits and 50 requests to AI-keywording. It’s totally free!
— On the Attribution plan, you will get a 30 GB file storage, 300 submits, 1000 requests to AI-keywording and the ability to connect two people to your account. It costs $6.99.
— On the Professional plan, you will get a 200 GB file storage, 1800 submits, unlimited requests to AI-keywording and the ability to connect two people to your account. It costs $12.99.
— On the Expert plan, you will get a 500 GB file storage, unlimited submits and requests to AI-keywording, plus the ability to create two projects and to connect three users to your account. It costs $34.99.
Try AI-keywording right now! Just sign up on dam.everypixel.com, it’s free :-)
-
Sounds like a recipe for a lot more poorly keyworded spammed files.
(I have not tested this keyword generation service, it might in fact be very good, BUT judging by the quality of almost every other supposedly great machine learning AI thing I suspect a lot of mistakes - that is fine if you understand the language you are keywording in and correct accordingly, but if you can do that you probably don't need this tool anyway).
-
Sounds like a recipe for a lot more poorly keyworded spammed files.
You can estimate the quality of AI-keywording on the demo page: labs.everypixel.com/api/demo
We trained our neural network to describe the objects on the dataset of keywords that commonly used on stock sites. That’s why it’s more relevance than results by similar services.
that is fine if you understand the language you are keywording in and correct accordingly, but if you can do that you probably don't need this tool anyway
Also, our tool will be helpful for native speakers too. According to our research, people usually pay attention to a smaller number of objects and their characteristics than a neural network. For example, somebody can forgive to describe the dominant colors or secondary persons. AI of Everypixel DAM will not forget :-)
-
Interesting idea - I know why everyone would love to automate keyword generation - but it really isn't up to the job. At least my experiments with several of my own stock images showed that in terms of keywording as well is the "stock photo score" the tool doesn't really know enough about what it's looking at to be useful.
I ran through a couple of kitchen remodel images and a couple of external shots of places where the important information was where it was in the world, not just what was in it. As these images have all been sold - in one case just shy of one thousand times across several agencies - I have an idea of what the important keywords actually are, primarily using SS's sales info that tells me the keywords often used for purchases.
In no case did the tool identify the place on the exterior shots and in the kitchen remodel cases, it missed all the key what-is-going-on elements and just picked up the filler - Residential Building, Home Interior, Architecture, Wood - Material...
In the case of one image, that was of an island ferry dock, it concluded that there was a Pipeline, a Factory, Construction Industry and Fuel and Power Generation - all totally incorrect. Plus it put the scene in Europe (it was off the west coast of the USA. Adding keywords like Nautical Vessel for boat may work with Getty's CV, but it's worse than useless elsewhere. No user types in these awkward terms and other sites don't translate them into the type of English real people speak.
Oh, and one of my really solid long-term sellers rated a stock photo score of 11.6%.
I think I tried enough different images to give the tool a fair evaluation, but I wouldn't use it.
-
Interesting idea - I know why everyone would love to automate keyword generation - but it really isn't up to the job. At least my experiments with several of my own stock images showed that in terms of keywording as well is the "stock photo score" the tool doesn't really know enough about what it's looking at to be useful.
I ran through a couple of kitchen remodel images and a couple of external shots of places where the important information was where it was in the world, not just what was in it. As these images have all been sold - in one case just shy of one thousand times across several agencies - I have an idea of what the important keywords actually are, primarily using SS's sales info that tells me the keywords often used for purchases.
In no case did the tool identify the place on the exterior shots and in the kitchen remodel cases, it missed all the key what-is-going-on elements and just picked up the filler - Residential Building, Home Interior, Architecture, Wood - Material...
In the case of one image, that was of an island ferry dock, it concluded that there was a Pipeline, a Factory, Construction Industry and Fuel and Power Generation - all totally incorrect. Plus it put the scene in Europe (it was off the west coast of the USA. Adding keywords like Nautical Vessel for boat may work with Getty's CV, but it's worse than useless elsewhere. No user types in these awkward terms and other sites don't translate them into the type of English real people speak.
Oh, and one of my really solid long-term sellers rated a stock photo score of 11.6%.
I think I tried enough different images to give the tool a fair evaluation, but I wouldn't use it.
Where you find "using SS's sales info" that show the keywords used for purchases
-
Interesting idea - I know why everyone would love to automate keyword generation - but it really isn't up to the job. At least my experiments with several of my own stock images showed that in terms of keywording as well is the "stock photo score" the tool doesn't really know enough about what it's looking at to be useful.
I ran through a couple of kitchen remodel images and a couple of external shots of places where the important information was where it was in the world, not just what was in it. As these images have all been sold - in one case just shy of one thousand times across several agencies - I have an idea of what the important keywords actually are, primarily using SS's sales info that tells me the keywords often used for purchases.
In no case did the tool identify the place on the exterior shots and in the kitchen remodel cases, it missed all the key what-is-going-on elements and just picked up the filler - Residential Building, Home Interior, Architecture, Wood - Material...
In the case of one image, that was of an island ferry dock, it concluded that there was a Pipeline, a Factory, Construction Industry and Fuel and Power Generation - all totally incorrect. Plus it put the scene in Europe (it was off the west coast of the USA. Adding keywords like Nautical Vessel for boat may work with Getty's CV, but it's worse than useless elsewhere. No user types in these awkward terms and other sites don't translate them into the type of English real people speak.
Oh, and one of my really solid long-term sellers rated a stock photo score of 11.6%.
I think I tried enough different images to give the tool a fair evaluation, but I wouldn't use it.
Where you find "using SS's sales info" that show the keywords used for purchases
Awww lookie they deleted their account LOL :D
Looks like the AI keyword tool is on the back burner then :D
-
Where you find "using SS's sales info" that show the keywords used for purchases
Well that was a short visit to the forum? ;)
Just in case someone else is new or has the same question.
Go to Top Performers: https://submit.shutterstock.com/earnings/top-performers?language=en
Only sold images appear, by earnings.
Downloads per Keyword: elvis (46.7%), presley (30.5%), memphis (8.6%)
click + and get more
elvis
46.7%
presley
30.5%
memphis
8.6%
tennessee
3.8%
stamp
3.8%
postage
1.9%
music
1.9%
50s
1.9%
famous
1%
Some images have none, some have what we'd expect and I don't know how valuable the information might be. In my case, if I see downloads on SS for the top few words, I sometimes adjust those words to first on AS. Since all my images have virtually the identical keywords, everywhere, I'm pretty sure, on all sites the "good" words are already in there.
Also for future images in my saved keywords, I include the most popular for sold, as the first, figuring it can't hurt to have them first everywhere, even though most sites, change the keyword order.
Lastly I can look at sold images and make sure that older images, with similar content, have the best words that were used to find the sold files.
Yes, there are many keyword AI tools already on many sites. Google says in the metadata, "looks like" and then what their AI thinks.
-
It is interesting to discover that behind Stock there is a big market of support services offered.
Similar as from simply Youtubing a walk to the zoo, it evolved to an industry
with so many aspects and services out of the actual simple "broadcast your self".
I always thought that Stock people do their tasks all by themself and total manually :)
-
Interesting idea - I know why everyone would love to automate keyword generation - but it really isn't up to the job. At least my experiments with several of my own stock images showed that in terms of keywording as well is the "stock photo score" the tool doesn't really know enough about what it's looking at to be useful.
I ran through a couple of kitchen remodel images and a couple of external shots of places where the important information was where it was in the world, not just what was in it. As these images have all been sold - in one case just shy of one thousand times across several agencies - I have an idea of what the important keywords actually are, primarily using SS's sales info that tells me the keywords often used for purchases.
In no case did the tool identify the place on the exterior shots and in the kitchen remodel cases, it missed all the key what-is-going-on elements and just picked up the filler - Residential Building, Home Interior, Architecture, Wood - Material...
In the case of one image, that was of an island ferry dock, it concluded that there was a Pipeline, a Factory, Construction Industry and Fuel and Power Generation - all totally incorrect. Plus it put the scene in Europe (it was off the west coast of the USA. Adding keywords like Nautical Vessel for boat may work with Getty's CV, but it's worse than useless elsewhere. No user types in these awkward terms and other sites don't translate them into the type of English real people speak.
Oh, and one of my really solid long-term sellers rated a stock photo score of 11.6%.
I think I tried enough different images to give the tool a fair evaluation, but I wouldn't use it.
AI-keywording it’s not human substitute, it’s a helper. The final decision about the content description still belongs to the person. Somebody can agree with proposed keywords; somebody can edit a couple of them; somebody can delete them all.
We admit that auto-keywords may be not enough relevance for you. That’s why our neural network continues to training so far. But an overwhelming majority of Everypixel DAM users satisfied with AI-keywords and slightly adjust them in some cases. Combined with the ability to upload content on several stock sites at the same time, it’s really perceptible time savings compared to manual tagging.
-
I guess if we keep training the computer by fixing what it comes up with eventually it will get better?
-
It is interesting to discover that behind Stock there is a big market of support services offered.
Similar as from simply Youtubing a walk to the zoo, it evolved to an industry
with so many aspects and services out of the actual simple "broadcast your self".
I always thought that Stock people do their tasks all by themself and total manually :)
The concept is as old as the hills
In a gold rush if you want to get rich don't dig for gold just sell spades
-
I guess if we keep training the computer by fixing what it comes up with eventually it will get better?
Yes, of course :-)
That's how AI works.
-
AI-keywording it’s not human substitute, it’s a helper. The final decision about the content description still belongs to the person. Somebody can agree with proposed keywords; somebody can edit a couple of them; somebody can delete them all.
We admit that auto-keywords may be not enough relevance for you. That’s why our neural network continues to training so far. But an overwhelming majority of Everypixel DAM users satisfied with AI-keywords and slightly adjust them in some cases. Combined with the ability to upload content on several stock sites at the same time, it’s really perceptible time savings compared to manual tagging.
Why do I want free storage? I can understand I want to upload, but after that, I don't need to keep my files on your server? What am I missing?
I'll assume anyone who has any business or uploads will want the $34.99 plan (why not just say $35? This isn't a Walmart) which is for how long? Expert plan doesn't have a time limit?
-
I guess if we keep training the computer by fixing what it comes up with eventually it will get better?
Yes, of course :-)
That's how AI works.
Call me when it can positively identify Greenbuls.
Other than that, it's easier and quicker to do it myself. But then English is near enough my native language; it would be different if I were trying to keyword in a different language. But even then, I'd need to know enough of the language to be able to delete false suggestions, so not that much of a help.
-
Why do I want free storage? I can understand I want to upload, but after that, I don't need to keep my files on your server? What am I missing?
There are two main use cases.
1. Some contributors (studios, most often) prefer to describe their content in advance and upload it to stock sites by degrees, according to the internal schedule. Of course, until that moment it’s more comfortable to keep files directly in service.
2. Many people usually keep their data in the cloud to avoid losing it. So, we give our users the possibility to save them content inside the Everypixel DAM storage after uploading on stock sites.
I'll assume anyone who has any business or uploads will want the $34.99 plan (why not just say $35? This isn't a Walmart) which is for how long? Expert plan doesn't have a time limit?
All plans are active for a month.
-
Call me when it can positively identify Greenbuls.
Other than that, it's easier and quicker to do it myself. But then English is near enough my native language; it would be different if I were trying to keyword in a different language. But even then, I'd need to know enough of the language to be able to delete false suggestions, so not that much of a help.
We are not sure that somebody can really easy and quickly to keyword by himself at least a couple dozen photos. On average, human needs more than 3 minutes to come up the ~50 keywords for one simple photo like a beach landscape or office routine. Our AI can describe it for seconds.
Now, imagine that you need to keyword 50, 100 or 200 photos in one run. It will be hard even for native English speakers.
-
Can your AI read existing metadata, for example description and generate keywords based on it too (+include words from description)? Or is it just visual?
-
Taking three minutes to keyword important words accurately is more efficient than getting 50 mostly inaccurate keywords in seconds. People actually want their images to be found and to sell.
-
Taking three minutes to keyword important words accurately is more efficient than getting 50 mostly inaccurate keywords in seconds. People actually want their images to be found and to sell.
I like using AI for faster keywording. It's good to get a lot of the simple words I sometimes forget but then I always take a step back and add my keywords too.
-
Call me when it can positively identify Greenbuls.
Other than that, it's easier and quicker to do it myself. But then English is near enough my native language; it would be different if I were trying to keyword in a different language. But even then, I'd need to know enough of the language to be able to delete false suggestions, so not that much of a help.
We are not sure that somebody can really easy and quickly to keyword by himself at least a couple dozen photos. On average, human needs more than 3 minutes to come up the ~50 keywords for one simple photo like a beach landscape or office routine. Our AI can describe it for seconds.
Now, imagine that you need to keyword 50, 100 or 200 photos in one run. It will be hard even for native English speakers.
It's the time taken to delete the 'false positives' that is the killer.
And I'd never submit that many photos at once.
Also, 50 words are seldom needed (sometimes, they are), and trying to come up with them just leads to irrelevances or spam.
50 words for a simple beach landscape, when you won't even be able to identify the location? That'll pretty certainly be spam.
Plus you'd then need to DA if if submitting to iStock.
-
At the point AI keywording is actually good enough to not need human handholding I figure all stock sites will be using it and we won't have to keyword at all. Sadly they will probably start using it well before that time and it will make the searches even more inaccurate and spammy (but maybe to counteract that the sites will push older properly keyworded proven seller files and our old stuff will continue to sell despite the tsunami of new and perhaps better content).
-
Can your AI read existing metadata, for example description and generate keywords based on it too (+include words from description)? Or is it just visual?
Just visual for now.
-
Just visual for now.
Just curious, is it technically challenging to combine both?
Because this way it would be much more accurate and since I'm sure the most important words are there anyway, then I may not need to waste time correcting the results.
Or is there hope that AI will be able to provide an accurate description in the future?
I doubt, since there are so many variables like location, abstract concepts etc. which can't be read by machine just because it will never be able to recognize that "this architectural detail on the building was shot in Amsterdam, not in Paris", so it will always need at least some data. Isn't it?
-
Just visual for now.
Just curious, is it technically challenging to combine both?
Because this way it would be much more accurate and since I'm sure the most important words are there anyway, then I may not need to waste time correcting the results.
It’s technically possible but not easy :-)
We are working on it. Probably, our AI will be able to analyze the metadata and generate the keywords at its base. But we can’t say specific dates.
-
OFFTOPIC
Perhaps if you want to attract people to the service, you might have to also focus the advertised text and on protection.
My first thoughts reading the thread was that there is no "security" reference, for the uploaded files, (I believe that exist, no offence)
also I wondered if random or selected copies of uploaded images exist forever as reference for the AI to learn and sample.
:)
-
My first thoughts reading the thread was that there is no "security" reference, for the uploaded files, (I believe that exist, no offence)
Why would you just 'believe' that? It was my first thought, and nothing has mitigated the thought.
Not saying it's an issue here, but I don't trust people just because they post, in this case pleasantly, on msg.
-
You are right ShadySue, trust online is not given easyto anyone.
but in a public forum, guess I have to keep my posts polite.
-
Why do I want free storage? I can understand I want to upload, but after that, I don't need to keep my files on your server? What am I missing?
There are two main use cases.
1. Some contributors (studios, most often) prefer to describe their content in advance and upload it to stock sites by degrees, according to the internal schedule. Of course, until that moment it’s more comfortable to keep files directly in service.
2. Many people usually keep their data in the cloud to avoid losing it. So, we give our users the possibility to save them content inside the Everypixel DAM storage after uploading on stock sites.
I'll assume anyone who has any business or uploads will want the $34.99 plan (why not just say $35? This isn't a Walmart) which is for how long? Expert plan doesn't have a time limit?
All plans are active for a month.
Thank you
-
I just tested it out. In my test of 2 images I think it is surprisingly decent. That is when compared to Google AutoML, Amazon Rekognition and Microsoft Azure. I've test the service of all 3 big tech firms within the last month for auto keywording. As others have pointed out, it is still not as good as a "knowledgable" human would do it. Back when stock photos was more profitable, like back in 2012, I use to send out keywording to be done in India. They speak English there and they wages were cheap, but I'd get questions all the time asking what things were in my photos. Sometimes they didn't understand the context because they didn't take the photo, and sometimes they didn't know what the objects were or didn't know the name in English.
Anyway, keep refining your software. It is only a matter of time before image recognition software is as good as a human.
-
I just tested it out. In my test of 2 images I think it is surprisingly decent. That is when compared to Google AutoML, Amazon Rekognition and Microsoft Azure. I've test the service of all 3 big tech firms within the last month for auto keywording. As others have pointed out, it is still not as good as a "knowledgable" human would do it. Back when stock photos was more profitable, like back in 2012, I use to send out keywording to be done in India. They speak English there and they wages were cheap, but I'd get questions all the time asking what things were in my photos. Sometimes they didn't understand the context because they didn't take the photo, and sometimes they didn't know what the objects were or didn't know the name in English.
Anyway, keep refining your software. It is only a matter of time before image recognition software is as good as a human.
More time than we would guess or what agencies and websites tend to promise.
(https://i.postimg.cc/zBZQKtT4/ss-why-ai-doesn-t-work.jpg)
These six are seen as similar. I had a rejection at another site for similar, when one was a BBQ chicken and the next image was a charcoal fire, similar color orange. :) I'm not ready to pay AI to do something that has to be cross checked and corrected. Yes I realize it's not perfect, but making almost as much work as saved keyword collections, I'm not ready yet.
-
Now, you can attribute files in a few clicks. Literally!
We have updated the Everypixel DAM interface to make the content attribution simpler and faster. Just upload your photos or videos 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.
The update affects not only the interface but also the neural network for auto keywording. We've expanded its video recognition capabilities: now it can find from 20 to 49 keywords, depending on the content and duration of the video.
Try it right now: dam.everypixel.com
(https://i.ibb.co/dWWYtnY/Sample-rus.jpg)
-
I can't create my test project because I always get this error:
-
Interesting idea - I know why everyone would love to automate keyword generation - but it really isn't up to the job. At least my experiments with several of my own stock images showed that in terms of keywording as well is the "stock photo score" the tool doesn't really know enough about what it's looking at to be useful.
I ran through a couple of kitchen remodel images and a couple of external shots of places where the important information was where it was in the world, not just what was in it. As these images have all been sold - in one case just shy of one thousand times across several agencies - I have an idea of what the important keywords actually are, primarily using SS's sales info that tells me the keywords often used for purchases.
In no case did the tool identify the place on the exterior shots and in the kitchen remodel cases, it missed all the key what-is-going-on elements and just picked up the filler - Residential Building, Home Interior, Architecture, Wood - Material...
In the case of one image, that was of an island ferry dock, it concluded that there was a Pipeline, a Factory, Construction Industry and Fuel and Power Generation - all totally incorrect. Plus it put the scene in Europe (it was off the west coast of the USA. Adding keywords like Nautical Vessel for boat may work with Getty's CV, but it's worse than useless elsewhere. No user types in these awkward terms and other sites don't translate them into the type of English real people speak.
Oh, and one of my really solid long-term sellers rated a stock photo score of 11.6%.
I think I tried enough different images to give the tool a fair evaluation, but I wouldn't use it.
LOL Joanne, I tried it with one of my top selling images from Adobe (high 100s not 1,000s of downloads). The number one keyword was "nautical vessel."
It's a small sailboat sailing near Newport, Rhode Island and the Newport Bridge, with a single passenger. Other keywords include "famous place," but no guess as to where it is. "Dusk" and "sunset" but not "sunrise," "blue" but not "orange," which is a very dominant color, and a few more iStock sounding phrases that I can't imagine anyone using in their search such as "mode of transport," and "man made structure." No mistakes, other than "yacht," but no mention of the man in the boat, no "boat," at all - there's "sailing ship," but not the number one keyword, "sailboat."
It got a 50.6% score. Not bad for a photo I took around 2009, which still sells regularly, but not like wildfire. It may be a fair assessment given its evergreen content. Or maybe it's ranked way too high. Joanne's should be much higher.
Interesting demo. I know I only tried one photo, but I can't imagine it helping me, though I don't really find keywording to be a chore, it's pretty simple. On the plus side, it could help non-English speakers (especially if they clear out the iStockspeak) and spark some ideas for newbies. It's more accurate than the AI suggestions that show up when you upload to certain stock sites, even if a few of the accurate words words and phrases are useless. And it doesn't seem spammy, which is a big plus.
IMHO, the real shortcomings here are the shortcomings I've noticed in all language-based AI that I've experienced - a lack of imagination - no concepts, and a surprising inability to pick out and identify the key feature in my photo using a word that real humans would use in their search: "sailboat" It's like Data, as brilliant as he is, tried his hand at keywording.
My guess is it will get better. But I think we experienced humans, with a good vocabulary and imagination, still have the edge.
To the OP, I hope these observations help.
However, as much as I'd like to see better, more accurate, and less spammy keywording, I'm not interested in paying for the privilege of uploading all my keywording expertise to help build the database, which seems like the real buy in here.
Decent job though. But that ship has sailed for me (couldn't resist) 8)
https://stock.adobe.com/stock-photo/id/85630579 (https://stock.adobe.com/stock-photo/id/85630579)
-
Just for giggles, I tried another one from Newport, Rhode Island. This time, an iconic lighthouse, to test its location knowledge. There is no other lighthouse in the world that looks like it and the surrounding cliffs are unique.
"Famous place" was the only marginally spammy phrase but it still couldn't ID the location.
And, again, it missed the most important keyword and the most important phrase - the word "lighthouse" and the name of the lighthouse. It's a close enough view to make out every stone in the building. RI is the smallest state - far fewer pix than of Paris - but Google can identify it in seconds.
But no spam - a real plus. And it liked this photo even better. 74.7%.
-
OK, I knew I shouldn't waste more time, but I tried an illustration, one with a few 100 downloads. It tends to sell a lot in February and March, since it's for Pi Day 3/14 (My daughter's a mathematician who works in AI - I designed it on a mug for her and eventually tested it out on stock).
Anyway, I'm crushed. It got a score of 0.00% and then, as if absolute failure wasn't enough, it told me, "This photo is very bad." (I nearly fell off my chair laughing so hard).
To be fair, I would give it a low score for saleability since it's such a niche illustration (not a photo) but it's among my top sellers, hitting 80 sales the first year it was online, and it's evergreen if quite nerdy, and niche, but reliable for its short shelf life each year.
I bet the coders would feel bad if they knew how mean their AI was to a mathematician's mother. 8)
(and they need to teach it about Pi Day!)
-
OK, I knew I shouldn't waste more time, but I tried an illustration, one with a few 100 downloads. It tends to sell a lot in February and March, since it's for Pi Day 3/14 (My daughter's a mathematician who works in AI - I designed it on a mug for her and eventually tested it out on stock). ...
our local pizza place always has a pi day offer - buy one, get the second for $3.14