About the talk
We study a Face Recognition problem caused by unforeseen in the training set variations in face images, related to artistic makeup and other occlusions. Existing Artificial Neural Networks (ANNs) have achieved a high recognition accuracy; however, in the presence of significant variations, they perform poorly. We introduce a new data set of face images with variable makeup, hairstyles and occlusions, named BookClub artistic makeup face data, and then examine the performance of the ANNs under different conditions. In our experiments, the recognition accuracy has decreased dramatically when the test images include an unseen type of makeup and occlusions, happened in a real-world scenario. We show that the fusion off the training set with several heavy makeups and other occlusion images can improve the performance. We show that the makeup and other occlusions can be used not only to disguise a person's identity from the ANN algorithms but also to spoof a wrong identification.
Stanislav Selitskiy is an Information Technology professional with almost 30 years experience doing research and development in a wide range of domains: from scientific numeric simulation to financial applications, from telecommunication and Internet service provider areas to machine learning in biometrics and NLP. He received his Master of Science degree in the Moscow State Technical University in 1991, majoring in Gaseous Dynamics and minoring in Software Engineering. He worked in research institutions, participated in Y2K projects, consulted in MCI/WorldCom, Sprint, AOL. He currently works as an anti-spam software architect in Earthlink Internet and collaborates with the University of Bedfordshire as a research assistant in Machine Learning studies.View the profile
Hello everyone and welcome back. We are glad you're with us and we are excited to bring our next presenter to the stage. We have Stanislaus. So this key research assistant and Senior software engineer from the University of bedfordshire taken away Stanislaus part of the world you are. So I'm going to present your work a little bit different perspective. Now, we'll be talking about note to you to show fakes. But we distinguish between real and fake one, but about how we find the same person with physical violence and makeup. Is acrylic
paint in Atlanta, but this research is done on the auspices of University of bedfordshire where I want you to research assistant is Don with help. And I want to thank all participants in this research on both sides of the Atlantic. And especially our grocery have a special That affected workers might be your girl. Ask why? We don't care about social cognition because in the last decade, you can say, the problem of correct above the human level face, recognition was so, so from the use a perspective, you can download. How old do you have to take a shower? Once you have your phone and you can
unload, cloak it using your face or if you are at a soldier. Again, if you're kind of, you don't even need to be part of the corporation, using the computer. If you could have just some game box with GPU, you can download library or like python or state-of-the-art air-beds, know what? That way to use on your train-set, replace your classes. And if you have a pleasant enough training set. Get effectively 100% accuracy for your test images. How old was Gio creation that remains? What if your test images. Representatives
in your what'll happen? For example, for real life, scenarios that could be your lighting relationship with emotion expression again or Mike. And Cindy said, he has been conducted since earlier. Order us four basic ignition in my chest and we still have some daughter said collected are they could be irrelevant right now? Either de montaigne, mm extended. Unfortunately. The abundance of propriety, or impropriety steelcraft number of Open Source data set of a clearance and
images of a kind of work will be done. On the doctor said, most of them collect at the game, like, in previous, Bluffs and minuses. We'll talk about this later. But what was electric company for a mold assessor, search. Unfortunately, it's not that clear-cut. So in some cases, those are my cops. Don't affect or evil needle stage one, but some of them. Do Portia real problems? Another interesting observation done by researchers of the targeted spoken as possible.
So you can masquerade one person and another and old days. I propose some solutions. Damn. You probably might want to switch to another ways of an educational face, but I raise. So if I do something but I like that. Logical. Alamance was it called. Before another problems with this research, again, internet is a great source of data about when you use old style fashion data, collection and controlled environment. With very structured, what I see of parameters are you might get something you won't get from going to go see your
own internet. You might get many subjects, but what I can say this object, Arabella controllable and not really big problem. I'm more simple Engineers features. So when we decided to take on this problem, we decided to collect our own that I said in the old style. And of course, it's publicly available. You can download it from Mandalay data and we try to cover as much as possible Variety in terms of age and race. I am so very best of different lighting. Conditions different emotional Expressions rotations. And also
interesting feature. We asked multiple artist for quail, pass of different styles and skill level. So what happened Christian day to hear state of the arts training for our number of subjects, verification randomly selected from all the other side? And of course, we'll go this way. Most 100, effectively, 100% accuracy. I bought the most interesting stuff off on began when we try to train or the train. Because we're used to transfer to learning methods. I will need to work on Mon makeup images and try to see the oldest, I could.
And my cups and against some of the Colossus was easily. Recognized almost no actually should drop by some of their miserable accuracy. Like a beetle not even more alarming. For some of these classes of my Carpenter Pluto's this state-of-the-art, not just miss identified. But Identify two ways about religion,. It's so to experiment with other controls. We, I wanted to test this hypothesis that saddle guidance will pick up what other needs with write. A disingenuous features bigger features. So, I just created a text to John space
detector. You come out and also another one. Connected. Network described in the front trying to get more Global. She chose to analyze and expectations of partially match. Yes, of course, it's obvious that some things which were amused by the coaches record by surprise. The opposite was also true. Tom closest Mitsubishi key songs by Alex lamp, surprisingly record by a vicious. So yes, it is. So fun symbol is working but we got another surprise open other kind from fully connected
PCA company in to CNN. The Tobias and I don't think so because she'd rather use a lot of low hemoglobin test you correctly. Identify makeup image or Image I usually and other one picked up 3-2, but interesting thing when one failed usually in other file, too, but with different candidate for, for the gifts again, something you might do not positive but negative way of against long. Could you get his medication and mother type of improvements in po practical suggestions? And what year was
right? Was training said if you were going so if we are Include into training said, some of the examples result might be not exactly the same. As in our testing. I'm just curious if it would improve performance and I will try to help a little bit because of me as a another funeral sound of waves, but beer. But if we had another type of a class, for example, is wide by done a bandana Alex note to self and 35 Min makeup cases. Erythrite, the effects of this Improvement, like basically
comparing distribution of the soft Max activation, which could be seen as probably has some cases. Exactly. It is compared distribution of wrongly, identified, images and images and how distribution move. Weather introduced another method to improve. Everything. Sherry slick example of accounting distributions like this, like one of his horn. This green one has correctly identified as wrong identified, wood worse. Sgm, take away from this. Maybe not exactly a way, but it's hard because the translation of trusted accuracy and trust a
probability against this image has a probability of Guinea. You will not use 90% or 99% of 75 for the decide. This is something that he's also done and we try to apply simple idea. Patient to see how important the publication's interesting thing that it's a really depends on their clearance over there application. Auto false positives, and false negative in the greater picture. Again. It depends on what kind of medication is example of Security application, where you trying to catch terrorists in the airport, you don't really care about four packages. And,
yes, you will talk to this person and destruction might be recognized as my way to the plane, but that was a true beer Station batting cages. For example, for legal application. You really don't want And what are the conclusions from our research before? In some cases, state-of-the-art current recognition, algorithms selection of old guy, using Whatsapp bicycle of images. If you have images of each note so I can always make up some story and what particular microphone on clearance at working best. Just a fool. You cannot beat old, goats
a spy glasses, fake beers, an examples of Makeup Kylie makeup contrast colors. An interesting point, not really sticky note anthropomorphic. So and a matter of professionalism. If you ask a really good printer to do the job because most of the amateur makeup given to work again. I'll find you. Some of my other person doesn't sweat but still in case in some cases we should really high reliability and accuracy points on Toronto Sun, supposed to be our location is file to form to report to the person you want. Assembles
different algorithms to help each other. So this is a proven way to improve overall system. Performance has a go for Infusion. And again, again, I apply to Portugal classes of images. If you see less than 90% at ignition. Probably. Yeah, you're about to throw away. This kind of I got a new song. Scott future work, of course, it's interesting to research other state-of-the-art Google. Now. I'm trying to do a little bit more interesting than the best. Little some ideas
down how to improve its Cascade Dish Network cell info. For example of a pacemaker future self. You can achieve, its officials recognized and used for the specification. Another topic is targeted support in and out of generators deepfakes. They could be a cretin print out, a map of the stock option to non-professional and maybe I want to see how it would work. Well, on this person to successfully. I'm here time for Christian suggestions commands. Hello, at 1, if you have any questions for stanislav, this
would be your time to enter him in the chat box for standing by. Thank you. Here is his question considering a used case of unlocked phone. So the goal is to identify particular face how much additional variance is typically explained by makeup features relative to use using purely geometrical information. For example, distance between pupils convexity of face Etc. This is, this is an interesting, actually. This is an interesting topic for research. We work with little bit and other creations. So yeah, this is open for grabs.
Question from a John a very 2020 questions. He practices. What's the overall performance on masked face has any interest in expanding the data set with more of these examples? Yes, actually, this is not old. We started making it since last year and it was originally Avada Kedavra, so this new virus it's going to be in East, Wareham masks masks and his daughter said, 304 examples. I would say that particular like bandanas, especially if this dark again and you should know small especially a small
details. They work better than life was and expiration and facial masks. OK. Google, order. Some kind of Macho musk, especially if you're interested in small and Veggie Tales, Adhikari songs by brother. Great. Any more questions, folks, please enter it in the chat. I think we may be at the end here stanislav folk. If you come up with any ideas and questions, please enter him in the chat throughout the day. This will all go into report and we will share it with all of our speakers. If something comes to mind after the fact, otherwise, we extend a great. Thank
you to Stanislaus and all that amazing information. And we'll have a thank you for joining from. Yankee girl. Then Joe now we will return at 9:50 where we have a ghost in my town and his presentation, another stretch break glass of water cup of coffee in a little snack and we'll see you at 9:50.
Buy this talk
Buy this video
Our other topics
With ConferenceCast.tv, you get access to our library of the world's best conference talks.