Member Discussion

Dahua Technology Sets New Record For LFW Facial Recognition

Here is a link to an article

Add it to the previous news on the Chinese companies and their interest to the latest artificial neural network solutions.

A year ago I started writing here at IPVM regarding the artificial neural networks, Deep Learning and facial recognition. One of the the first posts was:

This year there was a huge progress in a facial recognition - we have got solutions that are based on the Deep Learning. You can check current tests results (more than 30 developers) of scientific versions on the Labeled Faces in the Wild dataset. They are very impressive. And our own tests of a current and a new generation FR versions confirm it. Big angles, large and dark sunglasses, now it works with it. Of course, it will take time to create commercial versions. For now, these solutions are way too heavy. But I am sure that within next year we will see them on the market and within 1-3 years the current statement of Mr. Fernandes will not be funny. It won't be like on TV shows but it will be.

In many articles and discussions here at IPVM there is a message: video surveillance needs something new, something that will add value, something that will help to compete with Asian manufacturers, something that will push need in a high-end hardware and qualified installations. Biometrics, facial recognition in connection with the video analytics is probably the best way to make it.

It won't be easy and it won't be fast but anyway, we need to start working with data not only with a raw video. And it will be the worst mistake to ignore it.

It is sad, that one year later, we see Chinese manufacturers with teams and solutions and still arrogance and rejection from western manufacturers (WM).

Here are some ideas I have heard from WM in 2016:

  • It will never work - No comment
  • We will wait until it will mature - Innovations? No thanks.
  • Somebody will make SoC and we will use it - Somebody has made SoC but do not you think that Chinese products with the same SoC will still be cheaper? And SoC is a SoC - it has many limitations that are crucial when you try to apply it for complex and constantly changing operations.
  • Big deal, we will hire a couple of guys and they will make for us everything in-house - Good luck. Talk a couple of millions and years later when you start taking it seriously.

Asian manufacturers:

  • We have our own team working on it. / We partner with XYZ university research center in this area.
  • We are interested, let's discuss it. (Many of them do not have an enough budget for it but they are actively looking for options anyway.)

Artificial neural networks, it is much more than facial recognition or simple video analytics. A security system that can control security hardware, analyze a situation in a real-time and provide only valuable information to the operator, possibility to find any object or situation described in words, speech interaction with the system and so on. This is a future of using ANN. Will it be easy to make? - Sure, it will not be easy. Do not believe, want etc? - Just wait and you will be able to OEM it from Asian manufacturers (probably it won't be the tier one product but there will always be options).

Agree
Disagree
Informative
Unhelpful
Funny

Here is the test results summary from Dahua's original announcement:

The median score was ~.98, Dahua scored .001 higher than a bunch of other providers.

What does this mean in reality?

Agree
Disagree
Informative
Unhelpful
Funny

.001 is in fact a lot. I do not want to say that it will be the same in a real-life because most of participants focus on the result (i.e. very difficult to apply solution in real life due to the huge hardware requirements and slow processing) thus it is more interesting that they could make it better than other participants.

Most of all, it means that Dahua has a competitive ANN and a team that can create, develop and teach it.

Agree
Disagree
Informative
Unhelpful
Funny

.001 is in fact a lot.

Elaborate on that. It's fair to say that .001 advantage on many things is not significant. Why is .001 a lot here?

Agree
Disagree
Informative
Unhelpful
Funny

John, if you check the distribution of results it is pretty obvious that even .001 difference is a difference. For me it is enough to see that their result is .001 better than Baidu's result. Baidu as a main competitor of Google in China has made a lot (I am not going to go in detail but it was a lot) to reach this result and proof their superiority.

Anyway, it does not matter, in fact, whether they are the best or not. They are on the list.

Agree: 1
Disagree
Informative
Unhelpful
Funny

Why is .001 a lot here?

Its not. When the average margin of error is bigger than the difference, I don't think it is really significant. Looks like they should make the test harder.

On the other hand, its an encouraging result. How come no other surveillence manufacturers are interested in competing?

You would think Avigilon would be involved, if for no other reason but its a good place to round up a bunch of amnesty seeking video analytic manufacturers ;)

Agree: 1
Disagree
Informative
Unhelpful
Funny

Konstantin, is this a self-administered test on a known dataset?

Agree: 1
Disagree
Informative
Unhelpful
Funny

All information is here Labeled Faces in the Wild

Agree
Disagree
Informative
Unhelpful
Funny

I wonder what happened to this research back in 2010, this was ear detection.

Ears don't change shape in fact over age seems they change very little ( except perhaps the odd hair or two growing inside )

http://www.cspc.ecs.soton.ac.uk/ear

I found passport controllers in Asia check ears 80% of the time, if they are unsure about direct face. So there must be something in this.

Agree
Disagree
Informative
Unhelpful
Funny

In 2010 we did not have the same level of development in artificial neural networks as now. The break through was in spring 2015.

Iris is one of the most reliable modalities. But there is an obstacle it is difficult to capture a good image of iris on a distance. Because of this, facial recognition is more applicable for video surveillance and one of the first areas of ANN application. However, ANNs will be used for all biometrics modalities. It is a question of time.

Agree
Disagree
Informative
Unhelpful
Funny

From an established facial surveillance provider who is anonymous here as they do not want to criticize anyone specifically:

In general, algorithm performance tests are only valid if applied to unknown datasets, and if executed by an independent testing agent.

Agree
Disagree
Informative
Unhelpful
Funny

This is exactly what I was talking about.

This and similar tests are used by research teams from all over the world to assess and compare the performance of their solutions to others. Certainly, such kind of tests have known weaknesses but who cares? They have a team that capable to create and teach an artificial neural network and apply it to facial recognition. Do you think it is a piece of cake? And what will happen with established vendors (taking into account current prices) when they will push it to the market?

Agree
Disagree
Informative
Unhelpful
Funny

They have a team that capable to create and teach an artificial neural network and apply it to facial recognition. Do you think it is a piece of cake?

No, but it also does not mean that Dahua is anywhere close to releasing a production ready facial recognition offering. There's a lot more from passing a controlled test to dealing with actual field video surveillance feeds.

Agree
Disagree
Informative
Unhelpful
Funny

This and similar tests are used by research teams...

Pure research is one thing, commercial research with a view to issuing press releases trumpeting ones successes is another.

Since the dataset is known, and one can progessively tweak the algorithim to the specific dataset. This just happens naturally:

Programmer1: "Why is our program not recognizing image #268?"
Programmer2: "It gets confused because there is X and Y in this scene."
Programmer1: "Let's change it to handle that better, and try again"
Programmer2: "Ok, It recognized #268 but now image #353 is kicking out"
Programmer1: "Ok, Why?"

This re-iterative approach is useful to a point but it can be easily taken to an extreme. Where you are basically writing the program for the specific images.

Which is less likely to happen in pure research, but much more of a risk the more financially important the result is.

Agree
Disagree
Informative
Unhelpful
Funny

This re-iterative approach is useful to a point but it can be easily taken to an extreme. Where you are basically writing the program for the specific images.

Or colloquially called 'teaching to the test'

Of course, the question then becomes how good does the test represent reality. If it matches strongly, than great but otherwise false confidence.

Agree
Disagree
Informative
Unhelpful
Funny

Of course, the question then becomes how good does the test represent reality. If it matches strongly, than great but otherwise false confidence.

That's the problem here. We're not talking about a dataset type that represents a limited reality, i.e. straight-on without glasses, we're talking about specific images.

As an extreme example, one could write a program to 'recognize' someone based on nothing but the number of bytes in the image. Without even decoding.

Useless, but highly accurate.

Agree
Disagree
Informative
Unhelpful
Funny

Baidu, for instance, is not above cheating in such 'competitions'.

Agree
Disagree
Informative
Unhelpful
Funny

I'm curious to know how this stands up to NIST's FRVT as it appears not a single vendor from that test is here in this one especially well known FR vendors like NEC and Cognitec.

Agree
Disagree
Informative: 1
Unhelpful
Funny