"You need to invest in some studio lighting… The pic is not very interesting… It's flawed, but you'll get there: keep working on your lighting, framing, and composition. It's only a 3.2/10."
I think the fact that the AI is developed by Regaind (https://regaind.io/) must be taken into account. Their service is not intended for artsy pictures. It is rather targeted at people who have 1000 photos on their reflex from their last vacations and want to make a relevant and beautiful album of 50 photos to remember their trip, but don't want to spend two weeks comparing and selecting the right ones.
Exactly =) The target is not professional photographers but pretty much everyone else taking pictures everyday with their smartphones.
Artsy photos are more often than not wonderful for reasons which are not way too semantic for us for now - situations, etc, and which are way beyond what standard photographers do!
In my experience, even amongst vacation snapshots, the best tend to be of the accidentally artsy variety. I suppose this algorithm would completely fail to recognize the hilarity of the mirroring sphere of a tourist's perfectly bowling-ball-bald head in front of an apse, hovering under the blob of light that is the eye of the pantheon dome on a phone snap. I know I did, (miss it) on the first few times I scrolled through the series.
Generally I feel tempted to file this site under fun with phrase generators, with some ML thrown in to make it slightly more interesting than purely random phrase generator toys. But are there really no use cases for automated checks of aesthetic conventions? Sifting through the data dump of a full day of wedding shots maybe? I know I would be tempted to start with a computer generated shortlist. Mate it with face detection so that you can cluster by subset of attendees (each combination of faces captured must be present at least once in the final set) and... well I guess they already use systems like that, it's 2016.
An entirely unrelated idea in the general area of photography, algorithms and human learning that I still miss is this: an option in manual mode that takes a full auto shot right after the manual one, exposing how good or bad the algorithms are vs the operator. Or do some cameras have something like that? Do manufacturers omit it to not expose deficiencies in their algorithms? Do they omit it to not hurt the feelings of customers when they are worse than auto mode?
This is indeed a "fun" project which showcases the technology and provides (I think) fairly relevant comments (trust me - a random generator is much worse, we had terrible comments at the beginning of the project).
The other technology modules you describe (face recognition etc) are indeed requirements for a semi or fully automated curation system.
Indeed, it does not like my 4x5s, but it does like some anime figure pictures I took with my cell phone.
I think it basically looks to see if something is in focus, and you know the "rule of thirds", which it must have learned from custom-cropped photos uploaded to 500px.
wow, thanks for the link! those are really nice photographs, beautifully framed and very artistic (for pure art - that woman leaning on a wall, with the triangular shadow across the building, come on - wow!) and obviously even more informative to us 66 years later than when they were made - that is true art, wow.
one quick question - the page you linked starts with a top picture and then a set, by "the second picture of this fantastic set" did you mean the second picture on the page (gondolier in dark canal with tight ray of light from above) or second in the "set" (guy crossing tracks running with empty rickshaw)
Here I played with colors and color temp on a post-sunset landscape, and the resulting image, while really not great --in fact, plain weird-- is rated 9.3
Does the system have a notion of how sure it is of the rating it gives? [Edit: should it say "without a doubt"? There should be some doubt.]
(Also, is it deterministic (same image always get the same rating); it sure appears to be and it's certainly what people expect of a bot, but that's not how humans rate things...)
To be honest, I kinda like the image. It might not be great to you but it has a nice spooky fantasy air to it to me. As to wether it should be considered a good picture, I have no idea.
Hey there! I'm Regaind's CTO and will be happy to answer your questions :)
We built this as a coaching tool aimed at helping amateurs take better photos, in the context of building a smart and flexible automatic roll curation system that goes beyond the "bad exposition/motion blur detection" that's often used. The end goal is to fluidify the use of photos by killing the last bottleneck: sorting.
Don't get me wrong, I think this is pretty cool, but there is no real actionable feedback, so there unless you can guess as to what triggers the higher scores, there isn't much coaching, so much as "gaming" the AI.
My bias as an Asian American: I translate a 7.4 to a 74/100 to a C, which is not a good grade, hence a desire to see coaching to achieve a score of 9.8 and above.
Seems like a low score. Again though, it's not really clear why. "Bad lighting" seems odd, because maybe I have no taste but the lighting is the reason for the shot in the first place. For reference, here's the original: https://www.flickr.com/photos/stefandufresne/16091643487/in/...
So far, AFAICT, it is a big fan of single focus thing in foreground with lots of blur in the background.
Edit 2: Okay whaaaaaaatttt I'm really disagreeing with a lot here. Maybe I just think some of my shots are better than they really are, but 4.1 for this is crazy https://keegan.regaind.io/p/y3LaI_J0R6a33eJ_fOQfew
Algorithm seems to really like portraits. Every one I uploaded, it scored very highly. Threw in Afghan Girl by Steve McCurry https://keegan.regaind.io/p/yBJXfxHeTEWtvtGDxgupaA to see what would happen.
No face detection? It noticed there was a baby in my picture and commented on the expression:
https://keegan.regaind.io/p/XCBCz-9_TrC31c4Opa1mdQ
I'm sure there are many different definitions of face detection. Care to elaborate on what kind of face detection it doesn't have?
It does like contrast! Hm, the score change is sufficiently low to be within the error bars I'd say.
I agree that the upped clarity is less beautiful, but it adds a nice touch of sharpness which Keegan likes a lot as well!
When I have time later, I might feed it some images from the Magnum (the agency) corpus and see what feedback I get.
Once upon a time, in certain newbie photo critique forums, pranksters would upload a renowned photographers image for critique... Hilarious feedback from new photographers showing off their critiquing prowess.
I got it to give me a 9.7/10 but I seriously doubt I'm a better photographer than Ansel Adams I'm betting it has an unintentional bias against black and white photography.
But I get it. It's a great tool for beginners to see their rookie faults and address the basics [1]before they can foray into more intermediate photography. And then break the basics [2] in order to know why breaking the rules work and sometimes don't.
Semi-related (to the optics/mechanics) is this tool[3]
Truly impressive. I've been putting in pictures of my daughter from my phone and I think Keegan really shines at evaluating the pictures I take of my life. Sorta like a writing coach for a diary.
The difficulty will be understanding context. Was blurr intended for effect? Same with a fuzzy picture, or a dark one, or one with no apparent subject, or was it informational only, etc.
Some photography is presented as contemporary art, some isn't. "Fine art photography" is usually not credible contemporary art. It's often black and white, formalistic, maybe a female nude... I can imagine an AI that only understands formal qualities becoming pretty good at rating it.
What you might call conceptual photography, on the other hand, is "AI-complete". It's as hard as literature to appreciate, you definitely need a full human-equivalent AI.
Roland Barthes' ideas of the studium and the punctum start to open up the intellectual depth of that kind of photo.
This kind of project is possibly credible as a rater of formal quality against some ideal norm, maybe stretching to understanding composition in a crude sense. But everyone knows that it's the unexpected but visually compelling feature that makes an image good. (Or alternatively the significance of its content.) Standard strategies like the "rule of thirds" can only produce formulaic images.
I'm impressed with the ability to recognize different scenes and subject matters. There are achievements that include photographing waterfalls and beaches! How cool is this?
Side note: Is this bot on Facebook Messenger as well? Or did I misunderstand?
1) It utterly fails as a coaching tool. "Your framing is 71% correct". Great. Awesome. That's maybe technically correct (the best kind!), but it's utterly unhelpful.
2) It fails at images taken to capture a mood instead of a motif. (Exactly what I'd expect from ML :)
3) It has problems with things that we usually consider iconic photos. Annie Leibovitz's picture of Mick Jagger in an elevator? Meh, 5.2. Dorothea Lange's "Migrant Mother" gets at least a 7.6, so on par with my vacation selfies.
1) as I mentioned elsewhere, the explicit coaching happens when you submit a photo that Keegan think is bad (say score < 5).
2) Mostly true, except that it's surprisingly good at capturing the notion of good timing, which I did not expect from ML ^^
3) well, as for the Mick Jagger picture, I can see why Keegan dislikes it: it's simply not that good technically. What makes it great is that it's Mick Jagger and that he has a very deep sight (imho), which the model indeed does not capture. For the other picture, yeah, the score is not that good, but the comment is spot on (in my humble and very biased opinion) :p
Thanks for the feedback, and yeah, definitely needs more love before it's a perfect predictor! Nicely enough, classification tasks with tons of supervision still make a lot of errors too ^^ (http://pbs.twimg.com/media/CdOxQRbWAAEUZM6.jpg)
9.9/10 (https://keegan.regaind.io/p/9KfI67VdTGin64pOzN696g) - funny how a quick glance at the Hall of Fame and my brain can work out which of my pictures will match the preferred style in order to get a high score pretty readily.
I do wonder if it's a bit sycophantic, after all who wants to be told they're terrible all the time.
If you click the slightly hidden "get more feedback" button it shows:
Subject well framed: 100%
Super sharp subject: 71%
Great composition: 68%
Pleasant blur: 57%
Great lightning: 35%
Well chosen background: 33%
Interesting/original subject: 32%
Good timing: 26%
So it's not very happy with half the categories but scores 9.4/10, so the score is definitely not an unweighted mean! (geometric or arithmetic).
"Great color and amazing angle. It's okay, but I want to see more; keep working on your composition! A solid 5.7/10. Not bad, but I'm sure you can do better!"
How the heck is this feedback supposed to be useful / helpful / actionable?
That's a great example of the 'third element': composition, exposure and opportunity.
Sometimes you can make opportunity "happen" through equipment with more reach or foreknolwedge. But sometimes things just happen and a mediocre shot is more powerful than none at all.
Based on other people's attempts to please it, I thought it would like this close-up of a snail I took. 9.4/10 and a ribbon so I guess we know what it likes! https://keegan.regaind.io/p/rUMLzUc5SO-aZUoZzz31Tg
On the other hand, it doesn't like the lighting or colors on this, and wants more bokeh. I'm willing to grant that the composition is dull, but the colors? https://keegan.regaind.io/p/pQ9U6CiMRKWNkYVgqLZTCQ
And this one has the same light, and cleaner composition, but it still complains- I'm not sure what it thinks would improve it. "Bad composition" is not helpful advice here. https://keegan.regaind.io/p/6T6V3k8fRLSWMn2JT6-N4g
One of the issues is sometimes, it's okay to do things in a non-standard way for effect. The other issue is it seems to think the only thing that matters in photography is depth of field.
For example, this kind of generic picture of my son during the fall got an 8.2:
http://imgur.com/a/SOFE0
Where as this other picture I took, one I think is technically superior recently got a 7.1:
http://imgur.com/a/VeEL9
And then there was this one I took in Rome, which got a comment of "not enough blur" which is the exact opposite of what I wanted. This is a fiat in rome, the background is the entire point. rated at 6.2
http://imgur.com/a/zLStt
Compared to this pretty generic musuem pic... that's not really interesting in any way, but managed to get an 8.2 (probably because of the blur)
http://imgur.com/a/Szc9I
This is confusing as hell to me. I'm averaging 7.98/10 over 30 photos, which personally I think is too high. I'm not that good. But at the same time, my best photos are rated very low and my mediocre photos are rated very highly. This makes no sense.
This photo gets a 6.2/10 and is rated negatively for composition, but is actually near perfect (if I'd have a wider angle lens it might have been better). It follows the rule of 3s with reinforcing repetition throughout, including even in the perfectly captured reflections of me taking the photo itself. It's not an /interesting/ photo, but it's practically perfect composition. https://keegan.regaind.io/p/hrbE85wGQWW3vs5w1QoGFg
The scores range is explained in the modal, 5 is for decent pictures, 7-8 for great ones, 9 for exceptional ones. Your average is really good, and all the scores you got are pretty good as well, congratulations! Keegan definitely makes much worse mistakes than these ^^
Ribbons are one time events that have very different thresholds for happening, if you already got a given ribbon on one picture you won't get it again even if your new picture would be an even better fit for it!
Overall aesthetics is a very subjective subject, and getting perfectly accurate scores for it on the wide variety of photos out there is still an open challenge :)
Well these are mostly intended, but are kind of leftover of the phase where we had no user authentication implemented. You can still delete pictures from your profile (https://keegan.regaind.io/profile), that will remove them from Hall of Fame as well.
Your ribbon shows there by the way! (ribbons are implemented as one time events which depend on your user status, and are not stored alongside the picture. maybe we could fix that)
Thanks for the report tho, I'll see what we can do for reenabling the hall of fame button after the first show!
That's... strange. My guess is that it's saying this because of the overwhelming amount of green in the background. But taken as a whole this photo is almost a study in color. Clearly the algorithm discounts the artistic value of contrast lighting from the lamp and the differing colors of the rug and bookshelf contrasting with the greenery.
Posted a masterpiece by Stephen Shore, Keegan didn't care for it.
Static composition, my eyes don't really know where to look and the pic is not very exciting… It's flawed, but you'll get there: keep working on your composition and lighting. This is at most 3.8/10.
#messy#boring
I mean, is that a masterpiece? I ask as someone who is unaware of Stephen Shore and has never seen this picture before. As someone living in 2016, that looks like a random picture of a gas station someone took on a road trip. I agree with Keegan here, it's boring.
I'm sure there's some historical context or something I'm missing out on here, but that's just my take looking at the picture.
Stephen Shore is one of my favourites. It might help to think of this image as representative of a body of work, which makes it legible. (and, let's not forget, financially valuable). It's not the greatest Shore image.
Shore has a good "eye", which is something subjective that needs training to develop and recognize. It's very plausible that the particularity of a given artist's "eye" will be quantifiable by AI systems soon.
Shore wrote a popular book called "The Nature of Photographs". Recommended.
I am also unaware of Stephen Shore, and have not seen this picture before. I guess we can chalk it up to different tastes - I think this picture is amazing. It captures the feeling of the place.
It doesn't matter, the point is that a feeling is captured.
Many people seem to think that photography is solely documentary, but that is only one possible aspect. You also have control over what is in and out of the picture, where objects are relative to each other, the relative tones and brightness of objects, etc., and can use that to tell a story that's not strictly true. If you've ever been composing a shot, noticed that there is a big pile of garbage in the frame, and shifted the angle slightly, you're already doing this.
I have a DSLR and several fairly good lenses and like taking pictures, but I'm by all accounts a layman, and seeing as I'm unfamiliar with that photo, obviously not well versed in the industry.
But honest question: Is that a good picture? I don't assume that the AI is trained enough to give accurate critique on all photos, but in this case I might agree with it.
I'm not trolling nor mean any offense, but is there quantifiable skill behind that photo, or is it just as subjective as painted art? A simple example would be a Pollock painting which I don't care for at all.
Art MA holder here: Pollock is a really good example of the hard-to-pin-down quality of optical balance or energy or "eye". It's a learned thing like most aspects of culture. There are personal variations in "eye" judgement but enough consensus to keep the art market afloat. It does also have something to do with hierarchical structure, fractals etc., especially in Pollock's structurally simple drop paintings. His idea was a conceptual innovation but now does look pretty boring.
You might also dislike Pollock not only for being boring but for being associated with a tedious CIA-sponsored moment in art in the 50s, in which case I applaud you.
As amazing as the advancements to image recognition technology have been, I'm equally astonished of the progress corresponding language generation.
I would be interested to see a Turing test with this tool versus a real photographer analysis.
Restrict the "analysis" to 3 or 4 sentences along with a score. The participant goal would be to figure out which analysis came from the AI and which came from the photographer. Would be interesting to see the results...
I'm! You can get in touch with us at hello /at/ regaind.io
You can have a look at your webpage for all that we do (https://regaind.io), but yeah, automatic improvement is on the list of potential things we would like to do :)
>"You nailed the subject placement but that lighting is NOT working. Your portrait could be so much better. Try improving your lighting, background, and blur! No more than a 4.5/10, maybe less. Not so good, but I'm sure you can do better! Try again :)"
Don't know how to do those setting changes on my phone :-) I just posted a selfie on Bart... Bart lighting isn't too flattering
Well, the composition is not bad, the colors are more than decent. There's no bad motion blur, the main subject is clear (albeit gross). The score seems decent for that.
Actually, the score scaling is a very tough subject for a public thing. We received some comments saying that Keegan is wayy too harsh, it seems people expect their everyday picture to score 7-8 (we've been told it'd be a big problem in the US especially, I can't judge on that as not living there tho :/), while we feel that Keegan s generally a bit too generous for Europe standards.
I might be a particularly tough sell since I grew up in the US but have been living in Europe for quite some time now ;)
I think what you're seeing on is AI hype in action. The marketing of Keegan suggests that it's AI should be capturing some of what is usually perceived as subjective taste (e.g. whether or not the image is "interesting"). This ends up disappointing because the AI (which is clearly an incredible technical feat) is really just capturing objective elements of composition.
Perhaps you just need to fine-tune the language so as to do some expectation management?
Judge on a single sample, do you ? :) That's a fairly strong statement based on a fairly small dataset :(
But yeah, our heuristics to ignore some content (e.g. websites screenshots) sometimes trip on real pictures. There's still a lot of room for improvement :)
Yeah it's super harsh with street signs or that kind of large text on pictures. The protection is there to avoid saying really silly stuff on utilities bills and other kinds of paperwork. I'll add this example to the ones we should accept.
