pancakes

MicrostockGroup Sponsors


Author Topic: Shutterstock "Improves" Keywording With Autotagging for Mobile  (Read 5950 times)

0 Members and 2 Guests are viewing this topic.

« on: July 20, 2016, 14:59 »
+2
Shutterstock updated its mobile contributor apps yesterday and added a new keywording feature

http://www.shutterstock.com/blog/shutterstocks-autotagging-feature-for-mobile-makes-tagging-a-breeze

I don't upload iPhone photos, but I was curious about the keywording help - pretty doubtful that it could really do a good job, but wanted to see.

So I did an experiment this morning uploading an iPhone  photo of bear poop in my driveway to see how the upload process went. The photo was in focus, taken in flat, bright overcast light, but was pretty unremarkable.

The keywords SS's app came up with looked as if they came from an auto-spam generator! Beautiful? Sea? Red? Travel?

I entered 15 or so keywords, submitted it and within 4 minutes, I had both the "thank you for your submission" and rejection e-mails (lighting and composition). I did attach a model release for my foot - there to provide scale - so it wasn't rejected for that :)



So SS has only the 7 current images of bear scat/poop/feces - there may not be much of a market for the subject matter :) I'm not bothered by the rejection but

(a) this was clearly fully automated - no human looked at this picture and
(b) if you aren't a native English speaker, relying on this keywording tool is risky. It clearly doesn't understand what it's looking at

I did think about uploading a flower picture - like the blog example - but couldn't imagine they'd actually approve a cell phone flower (or food) picture just because of the subject matter.

YMMV


« Reply #1 on: July 20, 2016, 15:08 »
+4
Until somebody comes up with a computer program that actually "understands" what it's looking at, all "automatic keywording systems" are doomed to failure.


I reckon hope we're way off that time!


(Bear poop in the drive though! Scary!)

« Reply #2 on: July 20, 2016, 15:17 »
+1
LMAO, priceless !!!

the robot , guessing it is.. but it could still be a human considering the rejections many have been experiencing,...
cannot reject you for OOF, that's for sure.
it could have given you the boot for wrong WB
but robots do not really know if perharps polar bear poop could just well be blue in color.

the keywords ???
perharps had you included ---
-tasty
-delicious
-nutritious

atilla may have approved it  ;D ;D ;D

still, i think this is a pretty sh*tty topic

all in all, this sort of new-age artsy stuff is really not in big demand for an old-school stock agency like ss
« Last Edit: July 20, 2016, 15:39 by etudiante_rapide »

« Reply #3 on: July 20, 2016, 16:04 »
0
Architecture for a keyword? LOL!

Bear poop in your driveway...yikes!

« Reply #4 on: July 20, 2016, 20:01 »
0
i can understand the rejection for  composition,  but auto keywording is not a replacement for your own keywords but more asa tool to save you time or make things easier

« Reply #5 on: July 20, 2016, 20:01 »
+1
if no human looks at it, who detects trademark infringement

« Reply #6 on: July 20, 2016, 21:02 »
0
if no human looks at it, who detects trademark infringement


I'm guessing, but I would assume that the image analysis they say they've developed to support the reverse image search and keywording tools has some scoring system and if it sees something that looks like a logo or a face (with no model release and no designation of editorial) it gets flagged for human inspection.

http://venturebeat.com/2016/07/19/shutterstock-contributor-keywords/

I understand that the tool (as it's described in the above) will do better where the type of image is well represented in the collection - bear scat has 7 images out of 93+ million - but most of these keywords are completely wrong and none of the important keywords are there, so it really doesn't help much, especially if the goal is to assist people who aren't already great at doing keywords.

It's true that I can't know for certain that this was automated, but when you consider they're taking nearly a million images a week (meaning their intake is at least twice that), they're looking at 2 mill / 604,800 seconds, or over 3 images a second. They have to have some way of pre-screening or their inspection costs would bankrupt them.

With my serious submissions I send them in late in the evening my time and they're reviewed about 5 - 8 hours later. I can't see any likely explanation for the 4 minute turnaround on the iPhone submission other than the image failed the automated screening and didn't get a human review.

« Reply #7 on: July 20, 2016, 21:35 »
+1
Until somebody comes up with a computer program that actually "understands" what it's looking at, all "automatic keywording systems" are doomed to failure.

That's what it is supposed to be doing.  It's AI that analyzes the image and returns a set of possible keywords you can then choose from to add to the submission.  I tried it yesterday, and it at least got the idea that it was looking at food in a restaurant.  It's pretty cool if it keeps improving.  I'd rather tap a keyword to add it then type in the whole word.

« Reply #8 on: July 21, 2016, 00:12 »
0
how does it detect a design or trademark furniture or sunglasses which diesnt have logos or text?

« Reply #9 on: July 21, 2016, 01:20 »
0
Until somebody comes up with a computer program that actually "understands" what it's looking at, all "automatic keywording systems" are doomed to failure.

That's what it is supposed to be doing.  It's AI that analyzes the image and returns a set of possible keywords you can then choose from to add to the submission.  I tried it yesterday, and it at least got the idea that it was looking at food in a restaurant.  It's pretty cool if it keeps improving.  I'd rather tap a keyword to add it then type in the whole word.
I would imagine the more "normal" the subject the better it will get and over time it will improve. They are only suggestions after all.

« Reply #10 on: July 21, 2016, 04:13 »
0
Until somebody comes up with a computer program that actually "understands" what it's looking at, all "automatic keywording systems" are doomed to failure.

That's what it is supposed to be doing.  It's AI that analyzes the image and returns a set of possible keywords you can then choose from to add to the submission.  I tried it yesterday, and it at least got the idea that it was looking at food in a restaurant.  It's pretty cool if it keeps improving.  I'd rather tap a keyword to add it then type in the whole word.


If it can just tell roughly what it's looking at then I'm impressed! If the suggestions are at least reasonably relevant, then as you say it saves typing.


I was thinking about it last night, and I thought that things must have moved on some to enable the use of autonomous vehicles, which must have some idea of what they are looking at to work.
 

ShadySue

  • There is a crack in everything
« Reply #11 on: July 21, 2016, 04:53 »
+2
i can understand the rejection for  composition,  but auto keywording is not a replacement for your own keywords but more asa tool to save you time or make things easier
Not when you have to read them through and delete/uncheck all the rubbish. I'd much rather do it myself.

« Reply #12 on: July 21, 2016, 06:20 »
+1
i can understand the rejection for  composition,  but auto keywording is not a replacement for your own keywords but more asa tool to save you time or make things easier
Not when you have to read them through and delete/uncheck all the rubbish. I'd much rather do it myself.
That's the problem with it. Some people are going to use all the suggested keywords.
Keywording's one of "those" jobs though. It's not all that difficult really, but it's still a PIA.
There's also the perception by some that "more keywords is better". Nothing I've seen supports that, assuming the keywords actually describe what's there.   

Shelma1

  • stockcoalition.org
« Reply #13 on: July 21, 2016, 06:46 »
+2
Now that I think about it, my sales plunged when they introduced the new "similars" algorithm on the site, which looks for images with similar pixels rather than similar keywords. When that changed, the "same artist" and "similar images" that came up at the bottom of image pages changed substantially in many cases, and not for the better.

Before, with the keyword-based system, "same artist" would always show eight images in my port that were from the same series or at least the same subject matter, and "similar images" showed stuff a buyer might actually be interested in, like summer-themed images if they'd searched for "summer."

Now there are times when "same artist" will only show one image from my port instead of eight, and "similar images" shows images with similar colors, or overall patterns. For example, one of my summer images might have an orange burst in the background, and the eight similar images will all have orange bursts too, but the subject matter will have nothing to do with summer, which is what the buyer was interested in.

I know they're trying to be innovative and proprietary, but this pixel-based super magic algorithm is not working as well as they'd like it to be.
« Last Edit: July 21, 2016, 08:22 by Shelma1 »

« Reply #14 on: July 21, 2016, 07:22 »
+1
i can understand the rejection for  composition,  but auto keywording is not a replacement for your own keywords but more asa tool to save you time or make things easier
Not when you have to read them through and delete/uncheck all the rubbish. I'd much rather do it myself.

You don't have to uncheck or delete anything.  It puts the suggestions below the keyword field and you tap which you want to add.

Johnski2015

« Reply #15 on: July 21, 2016, 10:16 »
0
Does a bear s*it in the woods, no it s*its on Jo Ann Snover's driveway!

Great experiment though - made me chuckle - so the system basically doesn't work at all then - surprised - me not really

ShadySue

  • There is a crack in everything
« Reply #16 on: July 21, 2016, 10:33 »
+1
i can understand the rejection for  composition,  but auto keywording is not a replacement for your own keywords but more asa tool to save you time or make things easier
Not when you have to read them through and delete/uncheck all the rubbish. I'd much rather do it myself.

You don't have to uncheck or delete anything.  It puts the suggestions below the keyword field and you tap which you want to add.
And then add the importatn words which it didn't suggest. Still rather do it myself, but presumably if it's a common genre photo, it guesses better.


« Reply #17 on: July 21, 2016, 14:06 »
0
still saves you time, your suggestion was that it doesnt

« Reply #18 on: July 21, 2016, 14:21 »
+1
Now that I think about it, my sales plunged when they introduced the new "similars" algorithm on the site, which looks for images with similar pixels rather than similar keywords. When that changed, the "same artist" and "similar images" that came up at the bottom of image pages changed substantially in many cases, and not for the better.

Before, with the keyword-based system, "same artist" would always show eight images in my port that were from the same series or at least the same subject matter, and "similar images" showed stuff a buyer might actually be interested in, like summer-themed images if they'd searched for "summer."

Now there are times when "same artist" will only show one image from my port instead of eight, and "similar images" shows images with similar colors, or overall patterns. For example, one of my summer images might have an orange burst in the background, and the eight similar images will all have orange bursts too, but the subject matter will have nothing to do with summer, which is what the buyer was interested in.

I know they're trying to be innovative and proprietary, but this pixel-based super magic algorithm is not working as well as they'd like it to be.

+10
it's like everything in corporate affairs. there is the practical person who likes it to stay the same if it works well, ... as the saying, "the machine is not broken , don't fix it...
or else you only f**k it up !"
and then there is the F-rigging A-ggravating C-ontrarian T-weet(s) who insist on
showing off his IT know-how
and let FRED screwup take care of everything

so, we have ss being run by FRED
which only benefit a handful who keep insisting there is nothing wrong with ss these days

« Reply #19 on: July 21, 2016, 14:36 »
+1
I thought I'd experiment again - went on a field trip to the garden to take a flower picture with my iPhone 6s - to lob a softball at the SS mobile app to see if it did better than with the bear scat.

It did



I accepted many of those offered, added Azalea and Pacific Northwest (which it recognized after typing just the first few letters, so I just had to click to accept) and submitted. It took an hour for the rejection (based on e-mail times) - Focus and "Overuse" (noise reduction or sharpening). I assume it passed the first filter :)

This begs the question of who on earth would want to buy an image like this. The iPhone 6s has a pretty decent camera for a smartphone, but if you look at the out of focus background trees, the "bokeh" is horrible, and almost all iPhone pictures look over-sharpened to me.

Perhaps people have had iPhone images accepted to SS, but I have to assume lots would get technical rejections just based on the limitations of the camera.

However, even if you assume for a moment the technical details of the image were perfect, does improving this workflow really make any sense? If you were going to submit a flower/plant image, the identification of the flower/plant would be key - something the app can't do.

If you were to take this technology to the web submission process, it makes no sense as you'd be mad not to imbed your keywords into the file uploaded. So it'd be nice to have a keyword suggestion feature like this in Lightroom or Photoshop or a plugin to those apps.

Assuming they do want mobile submissions, I think the biggest help is their auto-complete as you type, which appears to be using their keyword database, not just iPhone standard stuff. Automated keyword suggestions could be improved:

1. Start with fewer and add more based on what the user picks up on.

2. Add a feature to clear their guesses and "guess again" (after you've typed a few)

3. Add some hints about what else the user needs to add - if the keyword picks out a human and guesses it's female, offer to add (one of) child, girl, woman, mature woman, senior (for example). It's easy to forget things you know - city, state, country; season of the year or morning/afternoon/evening/night

And adding a "did you forget" keyword feature to the web submission would be nice - their spell check has helped me out before when I misspelled something (and as Photoshop doesn't have spell check for the metadata I didn't notice). Yes, I have to go back to Photoshop and fix the original, but it's still better than propagating the mistake.

« Reply #20 on: July 21, 2016, 15:20 »
0
This begs the question of who on earth would want to buy an image like this. The iPhone 6s has a pretty decent camera for a smartphone, but if you look at the out of focus background trees, the "bokeh" is horrible, and almost all iPhone pictures look over-sharpened to me.

Perhaps people have had iPhone images accepted to SS, but I have to assume lots would get technical rejections just based on the limitations of the camera.

red i statement ... +10
red ii statement ...  not unless you are one of those new-age promoters of mob photos
                          with the notion that if ppl think selfies are so cool
                          they would definitely not care if the picture lacks clarity, WB,
                          and riddled by foreshortening,etc..

and assuming that ss will get more very satisfied contributors earning very little each month
because it cost nothing to produce tons of work for ss with their mobile.

then you get the accountability why older contributors are no longer needed
... you are all expandable!!!

not to forget, if all the people who have been making themselves redundant at ss H.O.
slacking off and eating pizza and letting FRED do everything,
when the time comes for ss to sack them all...
FRED would have already taken over the H.O.

more profit for the shareholders, and so sorry that rinderhart & compatriots will all be
standing in line at the gravy chain (greatest microstock photographers depression )  8)


 

Related Topics

  Subject / Started by Replies Last post
21 Replies
8154 Views
Last post October 16, 2008, 22:23
by madelaide
2 Replies
4016 Views
Last post February 18, 2012, 08:55
by dacasdo
10 Replies
5136 Views
Last post August 21, 2018, 14:55
by zorba
1 Replies
2820 Views
Last post May 11, 2019, 12:34
by davidbautista
1 Replies
2410 Views
Last post December 22, 2019, 19:12
by PhotoBomb

Sponsors

Mega Bundle of 5,900+ Professional Lightroom Presets

Microstock Poll Results

Sponsors