Luminous Landscape Forum
Raw & Post Processing, Printing => Adobe Lightroom Q&A => Topic started by: Chris Kern on April 21, 2015, 07:15:40 pm
-
Automatic face recognition isn't a feature I particularly care about — I've been manually keywording the people I want to be able to search on for years and it's not much of a burden — but I was really curious to see how well the Lightroom 6 (a.k.a., "CC") implementation worked. After an hour or so of playing with it:
- I'm really impressed by how fast it "learns." It only needs a couple of good exemplars to begin making accurate guesses about additional faces.
- The number of false positives is very low. This is important. You can tolerate misses (false negatives or regions that aren't recognized as a face), and in fact you should expect them, but misidentification of people requires a lot of manual checking to fix.
- As expected, faces in profile typically won't be recognized. As far as I know, this is typical of face-recognition software. You need to identify these faces manually, and Adobe provides a fairly intuitive method for doing that.
- I sometimes keyword people whose faces are barely visible, or aren't visible at all (e.g., back facing the camera), if their presence in the image may be a useful future search parameter. No way any face-recognition algorithm can do that; it's always going to be a manual process.
My only significant disappointment based on my initial experience is that once the identity of a face has been verified in one copy of an image, I wish that identification would be carried into other (virtual) copies. I think this should be possible, although there may be some constraint I'm not aware of.
All in all, I think Adobe did a pretty good job. While it may not be a feature of particular value to me (or many other Lu-La forum members, I suspect), I wouldn't be surprised if its presence will make Lightroom more attractive to casual snapshooters and, as a result, help expand the customer base.
-
Can you comment on the CPU overhead of having it enabled?
-
What is your hardware ?
-
It seems to take 100% of a single thread in the background.
-
Can you comment on the CPU overhead of having it enabled?
What is your hardware ?
I'm running a "Late 2013 Mac Pro" with a six-core 3.5 GHz Xeon microprocessor and 64 GB of memory. As jrp pointed out, the face recognition is a background process. Its presence isn't noticeable on this machine. However, Adobe apparently had some concerns about overhead because LR6 has a new process control panel, which opens when you click on the Lr logotype in the upper-left corner of the (primary) application window, that allows you to suspend the execution of background tasks.
-
Facial recognition was not on my wish list but I also gave it a try last night.
First I identified my wife on a folder of family photos, about 50 images and 20 or so with her face on. As Chris said, it learns quickly and after 5 or 6 manual entries LR was guessing quite accurately.
After that I moved to a serious challenge: A folder containing more than 1400 images of a parade with more than 8 faces (average) on each photo... including my wife on some. I started the process and went to dinner. One our later (I don't know exactly how long it took as I was away) it had identified 62 occurrences of her face and not a single one was wrong!
It is not very useful to me but it is quite accurate and fast.
-
Facial recognition was not on my wish list but I also gave it a try last night.
Mine either but I do like it a lot more now that I've tried it. It's also very entertaining to see what LR guesses as you add more face 'keywords' and it appears to learn. It even recognized some pictures of dogs and it's pretty funny when it mistakes people I know with a dog name or vise versa.
Anyway I'm working towards getting the entire catalog referenced this way. I made the mistake of picking All images. It bogs down big time. It appears to be more effective to work on a folder at a time, and from the bottom of the grid on up.
I was wrong about Face Detection based mainly on my own ignorance and not understanding how well Adobe would implement it into the Keyword pane. It will make searching far faster and easier.
-
The entertainment potential of Face Detection is almost worth the price of admission. Look what it found:
(http://digitaldog.net/files/FD.jpg)
Yes, it's a waterless urinal ;D
-
Yes, it is indeed entertaining. LR found faces in the hub of a power chair wheel, a mushroom, a luna moth on a tree, a hummingbird feeder hanging from a sheperds hook, some birds, some dogs, some fungus on a tree, in textures and in leaves.
Wonder what it would do with some images of Rorschach test cards? ;D
Jim
-
I wish there were some 'quality' index for filtering. Some images do contain faces that are rather blurry and as such, I'd prefer not to keyword them, even if I do recognize who they are.
-
The entertainment potential of Face Detection is almost worth the price of admission. Look what it found:
(http://digitaldog.net/files/FD.jpg)
Yes, it's a waterless urinal ;D
My first chuckle from the intertubes today. Thanks!
-
While it may not be a feature of particular value to me (or many other Lu-La forum members, I suspect), I wouldn't be surprised if its presence will make Lightroom more attractive to casual snapshooters and, as a result, help expand the customer base.
At first I considered this to be a useless gimmick.
But I suppose when dementia sets in and I am having difficulty remembering friends, family and faces I just may find a use for this.
-
At first I considered this to be a useless gimmick.
No more than keyword IMHO. And that's how LR treats them. If anyone uses Keywords and finds them useful for locating images, I suspect they will find the same true once the use the new feature. I'm not tagging all images found, but one's I feel are important to locate in the future.
I've learned a few tricks. If things bog down, you can quit and start where you left off, it seems to help. So you can search the entire catalog, produce the results in the grid, do some work, quit and restart where you left off later.
It's pretty interesting to see the product 'learn' as you go, grouping images correctly guessed with a ?. Select them and click on the little Accept icon.
If you have a group of images that are correctly guessed and a few that are not, select them all, click on the accept icon on one thumbnail that IS correct, the others take on that correct keyword.
Deleting images that are of no interest in tagging helps. The actual Delete Key seems a bit faster than the little reject icon, not sure why.
The "Sort By" option near the Film Strip is useful and toggling differing options can preset to you, images that are easier to select all at once.
-
Any idea how to change the default icon for a person in the People tab? It seems like the first image you tag of a person becomes their thumbnail icon for them as a person going forward. I'd like to be able to select a more flattering photo as their default icon, and also be able to change it as time goes on.
One thing I am impressed by is its ability to track with a person as they change in appearance. I've been applying it to my existing catalog with family photos, and it does a good job of keeping up with my kids' appearance as they grow up (my catalog goes back about a decade).
-
I think the any value of this will become apparent in time once we see how it deals with imports and if theer are any perfromance hits as a result of it's use.
I'm a little surprised that the feature seems unable to reference any existing keywords to make suggestions. A trick missed there.
I was also hoping for a more wide reaching recognition feature as someone here thought possible. Looking for predominate colours, shapes, features etc. Maybe that will added in future.
-
Not the feature I was missing but indeed, seems to work decently well.
-
The only other programme's FR feature I've tried is Picassa and LR6 is doing as well as that did.
Yes, lots of false positives, so some work needed to make sure it's recognising faces correctly.
What's unclear at the moment is, if, or how much, it learns as you add more positives to improve recognition of specific individuals. Also are there any specific ways to help it to learn ?
If it does learn with more samples, some effort put in when first setting the system up will pay dividends later in continuing use.
Difficult to judge, but I think it may have been more accurate on the last folders I ran through it compared to the first*, but it may be my familiarity with how to work faster with it now too.
*I've run 30k images through it so far.
-
I'm in the process of running facial recognition on my catalog of ~17K images. It took about 24 hours to complete the initial recognition processing on a 2010 MacBook Pro with an SSD. Now I'm in the process of confirming about 10K faces that LR couldn't positively identify.
One interesting aspect of this is that there are many familial mismatches so LR is in a sense able to tell which family members share similar facial features. (Of course my wife did not share my sense of amusement of LR thinking she was my son-in-law ;D).
The biggest issue I'm having is LR is just so frustratingly slow to update it's database when I confirm the name of face. It can easily take 10 seconds to confirm a single face, and if I attempt to group confirm, LR goes off for a minute and I'm just left with the spinning beachball.
I'm hoping that this will get better once I get all of the images tagged, but my fear is that LR just can't handle a large database of images and faces.
Scott
-
The biggest issue I'm having is LR is just so frustratingly slow to update it's database when I confirm the name of face. It can easily take 10 seconds to confirm a single face, and if I attempt to group confirm, LR goes off for a minute and I'm just left with the spinning beachball.
I didn't see that kind of slowdown. I did see operations slow down a bit and quitting then restarting seemed to help. My boot disk is SSD but the images and catalog all reside on an external drive, I'm working with a 2.7GHz i7 Macbook with 16GB of ram.
Maybe optimizing the catalog?
-
Trying to figure out how to 'force' detection of a folder after conducting FD in the first place. I tagged some but not all faces, now I want to go back and maybe do the rest.
Tried F1 (LR help) it takes me to Safari where no page is found. Ouch.
-
Just select the folder in the library pane, click on the little people icon and off it goes.
Check you haven't paused FD by clicking on the down arrow beside the id plate at the top.
No problems with accessing the help pages on Windows.
-
Just select the folder in the library pane, click on the little people icon and off it goes.
That isn't working. I've run this routine on the folder. I only selected some faces. Now I want to get the rest. It appears that once you've processed that folder, it 'think's' you're done. I want to force it to review the folder, exceed those faces I've detected and show me the one's it I didn't tag.
-
That isn't working. I've run this routine on the folder. I only selected some faces.
Ahh, it might not expect that behaviour from the user.
Now I want to get the rest. It appears that once you've processed that folder, it 'think's' you're done. I want to force it to review the folder, exceed those faces I've detected and show me the one's it I didn't tag.
Tricky. Nothing about that in the manual.
One thing that crossed my mind when going through the FR routine folder by folder was that there was no indication of what folders had been indexed.
-
One tip that seems to make a HUGE difference in how well LR identifies faces and how quickly you can tag people.
Instead of staying in the master view where multiple people are tagged at once, drop into a single person view by double clicking on the name/photo of a person listed in the "Named People" list. That will take you to a sub-view showing the Confirmed photos of that person, and below will be the Similar photos of the same person. Lightroom seems to have a much higher degree of accuracy finding similar photos as compared to attaching names to a wide swath of people. I've found that I can quickly tag hundreds of images using this method. I'll bounce from person to person, and the more I tag the better LR seems to get at finding similar images.
-
the more I tag the better LR seems to get at finding similar images.
It would be interesting and useful to hear from 'those that know' if LR really does learn and get better at identification with larger numbers of face confirmed.
-
After tagging 2678 faces in LR6 I can say yes, the feature gets more precise the more faces you confirm.
And I would also really love to know how to re-run the face detection process since I removed some face tags by accident and would like them to be re-detected. I only found these two following links:
https://forums.adobe.com/thread/1824626
https://forums.adobe.com/message/7471293#7471293
-
After tagging 2678 faces in LR6 I can say yes, the feature gets more precise the more faces you confirm.
And I would also really love to know how to re-run the face detection process since I removed some face tags by accident and would like them to be re-detected. I only found these two following links:
https://forums.adobe.com/thread/1824626
https://forums.adobe.com/message/7471293#7471293
There are many more posts about this from the UtoU forum, I've asked on at least three different posts there. So far, no answer from Adobe and it appears, I hate to say this, there is no way to run this again.
See: https://forums.adobe.com/thread/1824933?sr=inbox&ru=32475
-
There are many more posts about this from the UtoU forum, I've asked on at least three different posts there. So far, no answer from Adobe and it appears, I hate to say this, there is no way to run this again.
See: https://forums.adobe.com/thread/1824933?sr=inbox&ru=32475
Thank you, I have not found that one. For now I have backed up my LR6 catalog and reimported the most recent LR5 one so the scan could start from the beginning. Now let's just hope I don't make any mistakes until we get some reindexing option.
-
Northrup rates it around 60% vs Facbook is about 97%
https://www.youtube.com/watch?v=HLdgaV3rtcs&feature=player_detailpage#t=638
-
Faces was introduced in Aperture many years ago. I never used it because it takes up too much CPU.
-
I am reindexing ~20.000 images with ~2500 known faces in them, belonging to ~30-40 different people. If anyone would like me to do any test (not involving removing a face area from a known person) let me know.
-
One feature that should likely be incorporated into the next upgrade... LR should not suggest the same name (on a different face) in a photo in which that person is already tagged.
-
One feature that should likely be incorporated into the next upgrade... LR should not suggest the same name (on a different face) in a photo in which that person is already tagged.
Yes, I've been amused when it does that.
Jeremy
-
Wonder what it would do with some images of Rorschach test cards? ;D
Jim
Well lets start with it and find out the psychological identity of LR6 :-\ :-\