3 ms·

Sometimes I have my doubt on how current facial recognition technology gonna perform on tens of millions samples. Admittedly, we have become quite good at the s

by liuliu 16y ago

Sometimes I have my doubt on how current facial recognition technology gonna perform on tens of millions samples. Admittedly, we have become quite good at the scale of tens of thousands real-world data (see Labeled Face in Wild project), but how reliable it really is on large scale? My friend circle (everyone I have met in life for a small period of time) probably only consists of few thousands people, but I can remember many people who are distantly related have very similar faces. My main concern is, the facial variations between people may not be as large as we have imagined once we are on tens of millions scale. Here is an idea to validate my concern: 1. Crawl over all photos on Facebook and run a frontal-face detector with it; 2. Use current state-of-art methods (attribute based method or hybrid methods) to get several pairs of similar faces (highest score ones) of different people; 3. Use the same method as in step 2 to get several pairs of similar faces of same person with roughly the same score as in the step 2; 4. Have human volunteers to blindly judge which pairs of person are in fact the same one given only a small region of picture (e.g. face should occupy 20% of the total area); 5. Check the result in step 4 with ground truth data;

apu 16y ago
The face recognition problem is very tough and we are nowhere close to solving it for most real problems. The results on Labeled Faces in the Wild (LFW) [1] show that even on the easier "verification" problem -- "are these two images of the same person?" -- the best performance is under 90%. To figure out how this translates to recognition -- "who is this person?" -- you can roughly take the verification rate V and number of different people in your database N and get a recognition rate of roughly V^sqrt(N). (This is a very very rough estimate!) So with 1 person in your database, your rate is just V = ~90%. With 100 people, it's already down to 35%. And so on... (I'm the author of one of the best methods on LFW right now -- the attribute-based one [2].) [1] http://vis-www.cs.umass.edu/lfw/ http://vis-www.cs.umass.edu/lfw/ [2] http://www.cs.columbia.edu/CAVE/projects/faceverification/ http://www.cs.columbia.edu/CAVE/projects/faceverification/
- liuliu 16y ago
  I am eager to replicate your attribute method but haven't got much time to do so. Very excited to know that you are on Hacker News too. My impression is that "verification" problem is already a leap forward comparing to "identification" problem which consists of fewer people and many samples. The problem space is much interesting now than before with the astronomically number of digital photos available on the Internet (considering that for about 3 years, the "verification" result are improved from 69% to around 85%). But my main concern is that our human being may not be good at "large-scale" face recognition problem after all. If that were true, the fundamental of computer-based face recognition problem would be flawed.
  - apu 16y ago
    We have results in our face verification paper showing exactly how well people do on the LFW verification task. The answer is: very good, at 99.2%. However, we also show that much of this is coming from the background and context in the images in LFW. The fact that those images were originally gathered from Yahoo News means that many images are taken at the same event -- with the same background and same clothes, etc., and people can use this to make good guesses, even without seeing the person's face! I think the vision community is thus looking for how to exploit context effectively. There have been a number of interesting approaches recently, from looking at "things" vs "stuff", to using additional information such as transcripts and captions, etc. And of course, there are still lots of people looking at how to better solve the basic recognition problem itself.

by liuliu 16y ago

apu 16y ago

The face recognition problem is very tough and we are nowhere close to solving it for most real problems. The results on Labeled Faces in the Wild (LFW) [1] show that even on the easier "verification" problem -- "are these two images of the same person?" -- the best performance is under 90%. To figure out how this translates to recognition -- "who is this person?" -- you can roughly take the verification rate V and number of different people in your database N and get a recognition rate of roughly V^sqrt(N). (This is a very very rough estimate!) So with 1 person in your database, your rate is just V = ~90%. With 100 people, it's already down to 35%. And so on... (I'm the author of one of the best methods on LFW right now -- the attribute-based one [2].) [1] http://vis-www.cs.umass.edu/lfw/ http://vis-www.cs.umass.edu/lfw/ [2] http://www.cs.columbia.edu/CAVE/projects/faceverification/ http://www.cs.columbia.edu/CAVE/projects/faceverification/

liuliu 16y ago
I am eager to replicate your attribute method but haven't got much time to do so. Very excited to know that you are on Hacker News too. My impression is that "verification" problem is already a leap forward comparing to "identification" problem which consists of fewer people and many samples. The problem space is much interesting now than before with the astronomically number of digital photos available on the Internet (considering that for about 3 years, the "verification" result are improved from 69% to around 85%). But my main concern is that our human being may not be good at "large-scale" face recognition problem after all. If that were true, the fundamental of computer-based face recognition problem would be flawed.
- apu 16y ago
  We have results in our face verification paper showing exactly how well people do on the LFW verification task. The answer is: very good, at 99.2%. However, we also show that much of this is coming from the background and context in the images in LFW. The fact that those images were originally gathered from Yahoo News means that many images are taken at the same event -- with the same background and same clothes, etc., and people can use this to make good guesses, even without seeing the person's face! I think the vision community is thus looking for how to exploit context effectively. There have been a number of interesting approaches recently, from looking at "things" vs "stuff", to using additional information such as transcripts and captions, etc. And of course, there are still lots of people looking at how to better solve the basic recognition problem itself.