Giving Up on Picasa Face Recognition
NOTE: Images with an icon next to them have been artificially shrunk to better fit your screen; click the icon to restore them, in place, to their regular size.

I've just wasted the last two days on something more fun (and more frustrating) than any video game – Google's Picasa photo app “face recognition” stuff.

Google's Picasa desktop photo application includes the very cool feature of face recognition. Point it at a photo and it will identify where in the photo faces are to be found, and do so with great precision. I pointed Picasa at the 20,000 photos in my archive from this year (maybe 100,000 faces?) and there were no false positives, unless you don't count sculptures of faces as a face. Very impressive.

But like a modern Star Wars film (lots of cool effects, but no plot or acting or anything else that might contribute to making a good movie), the end result was disappointing because of a clumsy UI, poor recognition of who the face might be, and worst of all, a decidedly random approach to remembering what faces I've said belong to whom.

It took about a day and a half of crunching in the background for Picasa to chew on the 20,000 photos I threw it at, after which it grouped all the faces by person (as best it could figure), letting me ascribe names to the groups, and to fix errors in the groups. The application UI offers several different ways to approach the task of assigning names to faces and vice-versa, each with their pros and cons. Unfortunately, all the cons are severe. In one mode you can easily group photos so that the faces are all of the same few people, but it offers no way to group the results so you can indicate “these 20 faces are all such-and-such”; you actually have to actually type their name 20 times.

Another mode might allow you to indicate that a face should be ignored (e.g. some random stranger in the background of a shot), but other otherwise-more-useful modes don't offer that important feature, so if you see such a face while looking at suggestions for one person, the best you can do is indicate “not this person”, and so you're guaranteed to see the face again (and again and again) until you come across the face in the mode that actually allows you to ignore it.

Some modes make it very easy to say “these faces are/aren't this person”, but they make it very difficult to say “these faces are this other person.”

So, it becomes a pattern of jumping around modes to use each mode for what it seems best for, and you'd see the same faces over and over as you said “no, not this person either”, until you got fed up enough with it that you visited a mode where you could indicate who it was.

It was oddly addicting. It sucked me in like nothing else has since I made the mistake of giving Magic Pen a try a few years ago. I worked on the faces all day today, despite having promised myself a dozen different times that “I'll stop after tidying up this last person's faces.”

I should point out two very important caveats to my complaints above: 1) face recognition is new to Picasa (and to the world), so it's only natural that the UI will be a bit rough. It's a “version 1” application, so should be given time to mature before one can really expect a smooth UI. And 2), Google provides Picasa for free. Complainers should tread lightly.

Anyway, in the end I did abandon Picasa, and won't be returning to it any time soon. I'd certainly love to have all my photos “name tagged” (so to speak), but Picasa seems to have a very nasty habit of randomizing the data. I might mark 20 faces as being of the same someone, then return in five minutes to see that only one of those faces is there, but 47 other random faces are there as well, marked “confirmed”. This happened time and time again. At first I thought that I must simply be doing something wrong, but careful testing toward the end convinced me that it's just random, so no matter how cool the initial “where are the faces?” technology is, everything else as implemented in Picasa is not at all ready for prime time.

I built a plugin for Lightroom to suck over the name information from Picasa, but I'm now much less excited to release / maintain it (UPDATE: but nevertheless, I have released it here). It'll be interesting to see what Google does with this in future version — I don't doubt great things from them — but for now I'll just rely on my memory for photo face recognition. (It's also random, but the UI is better. 🙂 )


All 13 comments so far, oldest first...

I’ll say that Magic Pen is addicting… I just lost an hour trying to figure out the most Rube Goldberg way of kicking the ball up a teetering ramp….. and that was only level 2! I think I’ll quit before the withdrawal starts (reminds me of playing Tetris in my sleep 20 years ago….)

— comment by JasonP on December 8th, 2009 at 3:52am JST (8 years ago) comment permalink

At the risk of sounding obvious and maybe you mentioned this in the past but iPhoto 09?
Its on your new iMac and is supposed to do this superbly.
I tested it and it seemed very accurate.

— comment by Wayne on December 8th, 2009 at 4:03am JST (8 years ago) comment permalink

I haven’t seen the same, but honestly, I played with a much smaller set of photos. The only issue I personally noticed happy with baby photos (of my niece & nephew, when they were both under a year — somewhat reasonable).

The Picasa team is pretty solid. I’ll ping them to see if they know about the randomization issue.

— comment by Bill on December 8th, 2009 at 3:40pm JST (8 years ago) comment permalink

I was looking forward to your proposed use of Picasa for facial recognition and am disappointed to hear of the randomization problems. I shun iPhoto because of the way it takes command of the any photos showing up in iPhoto. Picasa seems to be kinder, as it only indexes the files for its own use.

Be careful with Picasa in this regard… I believe it does try to write data to the photos if it can. Prior to invoking Picasa, I made sure that all image files were read only, and that I had a full backup, just in case. —Jeffrey

— comment by ina on December 8th, 2009 at 6:36pm JST (8 years ago) comment permalink

My thought about a workflow for Picasa facial recognition was to download the files fresh out of the camera into a temporary folder in Picasa on a one-time only basis. I would then do the facial recognition and tag the folders and then right away mport them into Lightroom and read the metadata from the files. After that I would geoencode them and do whatever I normally do in ligthroom. At that time I would a;so delete the folder index from Picasa, so that the only access to the files would be through Lightroom and Photoshop.

In other words I would use Picasa ONLY for facial recognition on files fresh from the camera and before embedding any other info into the files.

Do you think this proposed workflow would work or are you suggesting that Picasa “corrupts” or the files in some way???

Some people prefer their image files to never be changed (the whole “non-destructive workflow” concept). In my case, any change (by any application, for any reason) to the contents of my Nikon NEF files destroys the image-authentication signature. (I don’t even know for sure whether Picasa would try to write them, but I didn’t want to take a chance). If you don’t have the “files are immutable” constraint , then it’s a nice feature if it writes the data directly to the file…. it’s certainly most convenient that way. —Jeffrey

— comment by ina on December 9th, 2009 at 7:31am JST (8 years ago) comment permalink

Very good point about the potential to destroy the image-authentification signature. In my case, however, I always save the original file (for safekeeping) to an external disk before doing anything to the copy.

In a “perfect world.” Adobe would be able to offer facial recognition built-in to Lightroom, and this wouldn’t be an issue. 🙂

Thanks,
Ina

— comment by ina on December 9th, 2009 at 9:42pm JST (8 years ago) comment permalink

I read your experiences with Picasa face recognition and decided to try a different approach.

I selected about 50 friends and family and selected about 500 pictures total in Lightroom that contained their images in different poses. I exported them as 600 x 800 96dpi JPG files to a new directory.

I set Picasa to work on them and identified the people whom I care about. As it was a small number of pictures with a few people Picasa processed them in less than 10 minutes.

It was pretty easy to establish names for my top 50 and then drag unknown pictures to those names. I did see one problem that it kept on showing me faces I am sure I told it to ignore. But apart from that the manual work in Picasa was quick.

By doing this I hoped to train Picasa to recognise the people I care most about. My next step when I have time is to point Picasa at all my photos. I hope it will have a high success rate identifying my Top 50 and I want to be able tell it to ignore the rest.

But I am worried that Picasa will confuse Lightroom by altering the master image files so I have not tried it yet.

What do you think of this approach?

Ian Fuller in Bangkok, Thailand.

Since I know nothing about how their face recognition works (I say “their” as if I know how any face recognition works 🙂 ), I don’t know whether the “training” actually matters, but it seems to make sense. But with this scheme, how do you intend to use the results? —Jeffrey

— comment by BKKPhotographer on December 11th, 2009 at 11:51pm JST (8 years ago) comment permalink

It was an experiment to see how best to use the very cool tool on a large number of photos. I don’t went to try and recognize every face in every photo I have taken. I already use Lightroom keywords to tag pictures of F&F so I don’t think I have a major use for it. But it seemed fun to try.

I bet a lot has been researched on face recognition technology. Are surveillance cameras common in Japan as they are in the UK? There are very few here in Thailand.

Surveillance cameras are common in stores and banks and such, though I’ve never noticed one in public like they have in the UK. Fifteen or so years ago I was wandering around one of the labs at the central research site of my company (Omron Tateishi Denki) and noticed some license-plate recognition stuff being tested. A camera had been mounted over the street, pointing mostly down, zoomed up fairly strongly to capture the bulk of one lane. As a car would flash by through the field of view, the system would isolate the license plate, decode it, and display the number faster than I could myself find the license plate. It was very impressive. (This system is used today to predict traffic times… you’ll see a big sign over the road telling you how long it will take to get to such and such a spot via what kind of route…. they keep track of license plates that passed the location, then passed various intermediate locations, then the destination location, and compute the times. Unbelievably, it is not used for speed enforcement.) A friend here in Kyoto (Britto) has worked on face recognition hardware for cameras… I’m surprised he hasn’t commented yet. I’ll have to ask him about it…. —Jeffrey

— comment by BKKPhotographer on December 12th, 2009 at 9:21pm JST (8 years ago) comment permalink

Hi Jeffrey,

Please read:
http://forums.adobe.com/thread/538646?tstart=0

I keep the information in XMP, not writing any information to the NEF/JPG files.

And if you are interested in joining forces on this project, get in touch with me. There are of course some issues of Picasa naming the ppl wrong, and that being inserted to the file.

I’ve looked into making it a true Lightroom plugin, but that will take some time – So for now its a 3 step process as described at the above link.

Best regards,
Michael

— comment by Michael Nielsen on December 15th, 2009 at 12:45am JST (8 years ago) comment permalink

Jeffrey,

I think I got along better with the UI than you, mainly by multi-selecting wrongly identified faces and then right click/moving them to the correct people. But it certainly shows all the signs of a first release…

Dan

— comment by Dan on December 24th, 2009 at 10:33pm JST (8 years ago) comment permalink

Jeffrey,

The other day I came across an article about Face.com that provides facial recognition as a service – and they provide a free service for up to 120,000 images per day (5,000 per hour). Facebook uses them as a provider in their backend. It seems like it may suit your needs if you are still interested in this sort of thing.

http://developers.face.com/docs/

Aaron
San Francisco, CA
Picture Site:http://www.rothytography.com

— comment by Aaron Rothschild on February 12th, 2011 at 1:39am JST (6 years, 10 months ago) comment permalink

Write-protection of files does not help with Picasa. Picasa v3.9 is disabling the write-protection of read-only files. Picasa adds metainformation and leaves then the files without read-only attributes. This is really a disaster! Other image software accepts read-only files. Some good DAM software even uses or honours XMP sidecar files in case of write-protected files. But only very few good DAM software is supporting XMP for read-only JPG files.

Regards from Rudy

— comment by Rudy on May 25th, 2012 at 11:20pm JST (5 years, 7 months ago) comment permalink

Picasa’s face recognition stuff has VASTLY improved since you made this post. I’m using it to people tag all my photos now (I like it a bit better than the Windows Photo Gallery app, personally). The only quirk I have now is in the “unnamed” group – it looks like it limits to 200 faces in that group at a time, and the times when it chooses to re-populate it from your unnamed faces is a bit odd. Rather than adding them at the end of the group, it adds them to the beginning, which means that the photo you were just about to tag with somebody’s face is no longer first in line, it’s 10th. When you’ve got a real flow going, it can be a bit distracting. Otherwise, the UI is pretty decent. It’s the only thing I use Picasa for though, I’m planning to tag all my photos, then import the photos into Lightroom using your plugin (hopefully your plugin still works with the latest Picasa and Lightroom).

— comment by Ryan Waddell on June 18th, 2012 at 11:16pm JST (5 years, 6 months ago) comment permalink
Leave a comment...


All comments are invisible to others until Jeffrey approves them.

Please mention what part of the world you're writing from, if you don't mind. It's always interesting to see where people are visiting from.


You can use basic HTML; be sure to close tags properly.

Subscribe without commenting