10

I am referring to the news item here.

Here's the whole issue, given a black and white video with mute sound, is it possible to tell that the actor 99.99% resembles a certain prominent figure? In the article it says that:

Shafee further said that according to the June 4 report by professors Hany Farid and Lorenzo Torresani from Dartmouth College there is a 99.99 percent match of the male in the video clip to a known photo of Anwar based on facial recognition analysis.

But to my layman judgement, it think it is quite amazing that one can readily identify someone to such a degree of precision in a black and white video.

So, how did the experts manage to arrive at the 99.99% conclusion?

Edit: From my research, it seems that Hany Farid is a pioneer in digital forensics, the detection of doctored photos. Matching a face in a video to a photograph doesn't seem like his expertise. A red flag.

Edit 2: It should be remarked that, the 99.99% remark is based on a secondary report from a guy name Shafee (not exactly the most impartial guy on earth, I must add), and so far Hany and Lorenzo have been silenced on this. How much can I trust a second-hand report from a biased party who has a political agenda behind him?

  • I wonder if there is a marketing interest hidden in there somewhere. – Randolf Richardson Jun 25 '11 at 7:08
  • 4
    @Randolf, nope, no marketing interest. You are looking at the wrong thing here. It's the political interest that is hidden in here somewhere – Graviton Jun 25 '11 at 8:15
  • Heheh! It's difficult to know which is worse. – Randolf Richardson Jun 25 '11 at 22:19
  • I'd say that the only thing computer recognition can confirm 99.99% is that it's seeing something. Everything else, YMMV. – Fake Name Jun 28 '11 at 3:58
  • i think what he ment was thet he just gave the picture of the actor(who act the part in the video) to match with the video rather then give the accused picture. Furthur more, he had been in jail for like 10 years. He should be younger when the video was made. It's just a claim from sombody who they claim 'expert' from some remote collage in the US. Its just a 'Collage'.I never heard of the collage before the news. Its not like some expert from Cambrage or Oxford. There are as many collage as there is stars in the sky in the US. Either way,I've read Heny Farid work and This Lorenzo guy. It turn – user3977 Jul 9 '11 at 23:16
9

A possible misinterpretation of the quote. A 99.99% match does not mean a 99.99% chance that this is him. For instance, if (a) the analysis checked 10,000 different things to determine a match and (b) this particular analysis found that 9,999 of those things did match then (c) we would have a 99.99% match.

Saying that there is a 99.99% chance that the face in the video belongs to a specific person would need to check how many different faces could have matched with the same type of analysis. If 100 people could have matched, then we have a 1% chance that this is the right face based on the analysis alone. (Other information could later narrow it down further.) Saying that the analysis is 99.99% chance this is the right person means that roughly 1 out of every 10,000 people could match in the same way. More gritty details below.


Using Bayes' law, we see that the probability that the matching man is the same man in the video is:

(The prior probability that the man was in the video *
 the probability that a video of him would give a 99.99% match to his face) /
(The prior probability that it was another man in the video *
 the probability that a video of another man would give a 99.99% match to his face)

To get an idea of the likelihood of this person being the one in the video, we'd have to look at the accuracy and recall of this face matching system (i.e. a ballpark estimate for probabilities 2 and 4 in the above formula).

On the other hand, to say that

This means that A and B is the same person

Is not an equivalent statement without additional justification. However, it seems to be from a different source.

  • 2
    (The prior probability that the man was in the video * the probability that a video of him would give a 99.99% match to his face)/(The prior probability that it was another man in the video * the probability that a video of another man would give a 99.99% match to his face)-- Not to sure what you mean here, can you explain more? – Graviton Jun 25 '11 at 8:08
  • 2
    @Borror0: But you must not have understood my answer if you think that the number 99.99% enters into the picture as a probability. The only calculation I present is just a simple statement of Bayes' law. – Jeremy Salwen Jun 25 '11 at 21:21
  • 2
    @jer: you are saying that this is a misinterpretation of the quote and explain what it means in your opinion. The fact is, your opinion is not nearly enough here. Does an official, reputable source back up your interpretation? Then link it. – Sklivvz Jun 25 '11 at 21:31
  • 3
    I'm not offering any novel explanation. Do I need to reference a dictionary to establish that "99.99% match" does not mean the same thing as "0.01% chance that there would be a match this close and yet it still not be the same person"? – Jeremy Salwen Jun 25 '11 at 22:24
  • 5
    NO. You must provide evidence that Hany Farid and Lorenzo Torresani from Dartmouth College used 99.99% in your sense if and when they used that terminology. Surely, as academics, they will have explained exactly and contextually what they meant, in which case: find the relevant papers and summarize them here. – Sklivvz Jun 25 '11 at 22:36
0

The confidence of the match depends on the quality of the source, and the algorithm, of course. With video there is more information to mine; multiple angles of the target, the gait, and so on. Farid is a published researcher on forensics so he is qualified to make statements on that, not so much on identification, which is what we are discussing. Moreover, when I hear a nice round number like 99.99% I think the person is not being literal.

  • Welcome to our site. We expect ansewers to be qualified with references. – Sklivvz Aug 2 '11 at 6:01
  • Uhm, not very convincing. Being published doesn't mean much, By itself. Does he make the same arguments and statements in some of his published, peer-reviewed research? I don't think his status is relevant. – Sklivvz Aug 2 '11 at 6:59
  • Statements qualified. – Emre Aug 2 '11 at 7:07

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .