yssirk123 Loc: New Jersey

I saw Everypixel.com mentioned in some of the posts here on UHH, and thought I would see how their "neural network" would rate some of my pictures. All you need do is drag an image onto their site, and purportedly their AI algoritm will analyze the image and tell you how good it is. I'm not sure utimately how accurate their ratings are, but it seemed able to broadly distinguish between mediocre and better images. With the advent of mass market stock photography, I think services like this are a sign of things to come.

Also, it has a helpful keyword generator that you can use elsewhere. These are the keywords it generated from my picture which seemed pretty accurate:

Sea, Beach, Sunset, Nature, Coastline, Wave, Sky, Scenics, Sand, Outdoors, Water, Landscape, Summer, Rock - Object, Blue, Cloud - Sky, Dusk, Beauty In Nature, Sunlight, Sun, Everypixel

Overall I think the site is fun to use - what do you think?

Bill

(Download)

RobbieAB Loc: UK

Is that rating out of 100?

My reaction to the image posted is "another seascape sunrise/set, pretty colors, all nicely by theory, but it's still yet another seascape".

I'm not trying to be rude, it is a nice image, but it's really not standing out from the vast horde of nice seascape images we see.

selmslie Loc: Fernandina Beach, FL, USA

It's still at the "fun to use" stage, in Beta testing and under development. It is uncanny in its ability to pick up on content (keywords) but they can also be a bit off.

If you use it for a number of pictures you might find that it has a preference for people, animals, strong colors, contrast and B&W.

It still has a serious problem with some color images in that it reports a score of 0.0% or other very low value for images that, when converted to B&W, score much higher - even into the high 90s. I have reported this to them and provided them with some sampled.

Overall it does a pretty good job of ranking images but don't let some of the very high scores go to your head. Just as it should not be reporting 0.0%, some of the 99 to 100% scores are probably also wrong.

selmslie Loc: Fernandina Beach, FL, USA

For comparison, I took downloaded your image and when I converted it to B&W the score dropped to 78.4%. I actually like the B&W version better.

Because of the problem they seem to be having with the extreme scores you should probably take them with a grain of salt. You might need to look at the scores for the other version when one of them comes too close to 0.0% or 100.0%.

I also took one of my low-scoring images that I knew was too dark and just increased the brightness. The score went from 1.9% to 15.0%. That change in score actually makes sense.

yssirk123 Loc: New Jersey

Thanks selsmie. I like the automatic key word generator too. I'm trying to figure out what its biases are by submitting different pictures; it definitely likes bolder colors. I also found that if you have an image that scores 0.0, just refresh the page and let it re-score it.

yssirk123 Loc: New Jersey

Thats interesting - I think I will convert some pictures to b&w and see what happens to the score. I did find a number of middling images that scored in the 50-60 range.

yssirk123 Loc: New Jersey

Hi RobbieAB - thanks for looking.

DirtFarmer Loc: Escaped from the NYC area, back to MA

yssirk123 Loc: New Jersey

Hi DirtFarmer. Looks like file size is one of the parameters they use. BTW, thats a nice shot you used in the other post.

selmslie Loc: Fernandina Beach, FL, USA

If you send them those two files they will appreciate it.

It's probably a bug.

rgrenaderphoto Loc: Hollywood, CA

I question the ratings algorithm, but what is very useful is the keyword generator. For anybody who is thinking about an on-line protfolio with an e-store, one sure fire way to drive traffic to a site is keywords.

selmslie Loc: Fernandina Beach, FL, USA

The ratings algorithm is probably more objective and informed than the average critique we see here.

yssirk123 Loc: New Jersey

I agree on the keyword generator - seems like it does a pretty good job.

DirtFarmer Loc: Escaped from the NYC area, back to MA

It does get a lot of things right, but it needs polishing. In my example above it gets 7 of the keywords wrong out of 19. I'm counting the following keywords as wrong:
People
Men
Women
Couple - Relationship
Heterosexual Couple
Females
Togetherness

I'm counting them wrong because they are all plural. There's only one person in the photo.

Of course that's only one example and other images I've tried do pretty well with the keywords. The most common error I've seen (in a very limited sample) is counting things as plural when they are singular (mostly people). It seems to occasionally make unwarranted assumptions about male/female (or it seems to list men/women more often than singular). I will say that it could be very useful to present a number of keywords that you could choose from. However it isn't yet at the point where it can generate the keywords automatically without human oversight. (In particular, I like to put names in my keyword list and I might worry if it got all the names right since that implies an immense facial database).

Also in the example I gave, I took the 0% one and changed only the filename and it came out 100%.

Peterff Loc: O'er The Hills and Far Away, in Themyscira.

Be very careful with AI and deep learning. An AI or deep learning system takes a very large amount of training to produce reasonably accurate results, so you may be part of the training program.

Remember a while back when Google photos tagged a picture erroneously and also in a completely unacceptable manner? Google is one of the leaders in AI, but the tech is still in its infancy except within very bounded circumstances.

Please look at the images below as an example of an AI / deep learning technology epic fail, not in a political or racist context.