Hats off to people like him who are helping to make computers understand humans. Facial recognition and speech recognition are two gigantic hurdles, followed by understanding language. I am amazed that we've come so far.
+D347Hza for 8 bits there exists 255 possible LBP patterns, by counting how many of each LBP pattern we see in an image we can predict what it might be.
I'm currently studying the suject for a school project, in fact there's 58 uniform patterns (with only 2 or less transitions in the bit sequence) and 256-59 = 198 non uniform patterns ("00010000" is uniform "011100001" isn't) but in the experiment you observe that 90% of obtained patterns are uniforms rather that 22.6% so to encode the image, you create a 59 dimension histogram and in the first 58 dimensions you count the number of occurence of each individual uniform patterns whereas in the 59th you count the number of every non uniform patterns. Then you do this histograms for regions of the image (containing multiple 3x3 cells) and you concatenate (put together but not merge) every histograms and you get the full picture ^^ To predict you make the weighted average of the previous positives histograms and you compare with the new one and there u go.
The example of 3D computation of LBP is interesting, using adjacent frames. But I ask myself if we can use this method in coloured image RGB? Especially to recognize coloured texture for get the variation between colour. Or maybe this kind of approach is different.
This guy looks exactly like my friend. It's not an "oh yeah, he does look a bit similar" type thing, it's genuinely as if they could be twins separated at birth. It's pretty spooky to be honest.
thank you for your explanation, is that possible that we create lbp algorithm (code) for face recognition from zero (without any libraries or modules)? because it will be easier if I can learn from the source code
So what your getting at here is that doing it in the last way mentioned increases the speed that it calculates the data but there has to be at least one advantage in calculating it at 2^26 is there not? Also does each pixel undergo this calculation or is it just every 9 pixels? I may just not grasp the full concept, Still not as tech savvy as I could be.
Why not say....remove the 8 edges of the cube which are less time/space relevant in terms of information in relation to the central pixel? Much more simple...
***** Yes, that is true. Alone, the counting rule applies to the data, with no compressions possible accept for one to one mapping with a few small values.
Please make a video about self driving cars. There are a lot of conflicting opinions about whether or not we should allow computers to operate a motorized vehicle on public roads, and there has been some news of cars being hacked. As a student of Computer Science, it seems to me that if done correctly it should be safe, all the same I'd like to know the opinions of others as well.
Ok for getting into the thinking of image processing but that technique is nowhere robust enough to handle the real world. The short answer is neural nets.
+Ethan uhm.. it reduces the search space from 2^26 to 3*2^8, the number of blocks encoded was not the point there. (just for for those curious, 3*2^8 is 768, 2^26 is 67.1 billion, quite the reduction)
+Ethan it does go down to 8, because it only considers dimensions 2 at a time. Each pair has 2^8 possible transitions, and 3 dimensions can be paired in 3 ways (XY, Xt, Yt), so this orthogonal method has 3*2^8 possible values. Don't be tricked by the image, the method is not just taking away the edges
Even though we've achieved amazing progress on object recognition in computing, we're not even coming close to a viable solution about it. I'm truly convinced that mathematics is NOT the answer here, as is not a good answer for every other NP problem. But, the video is great to show the very reason why - at any point, the algorithm is "aware" of only one single dot of reality. And a pixel is meaningless in the real world.
+Peter Walker I don't know :) If I knew, I'd probably have several Nobel prizes on my chest. Maybe some magic branch of mathematics' paradigms and models we haven't yet dreamed of. What I'm saying that the current mathematical and computing models of "real world awareness" is totally useless. Well, not totally, but far, far away from anything close to real-time.
I like the episodes like this that go into detail about an obscure topic without making things overly complicated.
A rare occurrence, where I have an understanding of a topic and still agree with the video. Very nice explanation.
Hats off to people like him who are helping to make computers understand humans. Facial recognition and speech recognition are two gigantic hurdles, followed by understanding language. I am amazed that we've come so far.
you have no idea!
4:51: 288 in 8 bit? Isn't 255 the maximum.
"You'd think i'd be better at drawing faces after 10 years, but i'm not"
Speaking of faces, Dr. Michel Valstar has a nice-looking one!
Thanks, now I've got a great method to figure out my girlfriend's moods.
pro tip on relations: likability is inversely proportionate to perceived words spoken.
Wow i love the animations, i think it really helped me understand LBP better. Thank you !!
4:50 288? I thought those were bytes, which would mean it could only go up to 255?
Probably misspoke, even if it's a more casual style interview, knowing you're on camera will often turn a brain to mush.
I would love videos which go into how. For example, show them writing code in whatever language they use to establish these things.
nice seeing LBP explained here. part of my thesis was working with LBP. it's a cool method
i am also working on the same. i need help in implementation of lpb on an image. can you help?
I dont think he explains how the numbers generated in each block get used. Do they dissect it again and map the edges?
+D347Hza have a look at the extra bits - this mentions what they do with the data obtained
+D347Hza for 8 bits there exists 255 possible LBP patterns, by counting how many of each LBP pattern we see in an image we can predict what it might be.
I'm currently studying the suject for a school project, in fact there's 58 uniform patterns (with only 2 or less transitions in the bit sequence) and 256-59 = 198 non uniform patterns ("00010000" is uniform "011100001" isn't) but in the experiment you observe that 90% of obtained patterns are uniforms rather that 22.6% so to encode the image, you create a 59 dimension histogram and in the first 58 dimensions you count the number of occurence of each individual uniform patterns whereas in the 59th you count the number of every non uniform patterns. Then you do this histograms for regions of the image (containing multiple 3x3 cells) and you concatenate (put together but not merge) every histograms and you get the full picture ^^
To predict you make the weighted average of the previous positives histograms and you compare with the new one and there u go.
Thanks a lot for sharing! I am complete new in this, I didn't understand, how to get it 288 in a pixel at 4:48
Excellent video. Very nicely explained.
The example of 3D computation of LBP is interesting, using adjacent frames.
But I ask myself if we can use this method in coloured image RGB? Especially to recognize coloured texture for get the variation between colour. Or maybe this kind of approach is different.
This guy looks exactly like my friend. It's not an "oh yeah, he does look a bit similar" type thing, it's genuinely as if they could be twins separated at birth. It's pretty spooky to be honest.
+sporkafife hah...try face detector on them, see if they match
thank you for your explanation, is that possible that we create lbp algorithm (code) for face recognition from zero (without any libraries or modules)? because it will be easier if I can learn from the source code
what in case when in pixel of coords [0,0] we don't have value of [-1, -1], [-1, 0] etc.?
This is a nice question, I am waiting for someone to reply it
In general, these algorithms performs a zero padding under the hood, but sometimes the neighbors can be mirrored/reflected to the padding
So what your getting at here is that doing it in the last way mentioned increases the speed that it calculates the data but there has to be at least one advantage in calculating it at 2^26 is there not? Also does each pixel undergo this calculation or is it just every 9 pixels? I may just not grasp the full concept, Still not as tech savvy as I could be.
Will LBP have some effect if I want to use CNN to image classification applications like say emotion recognition.
Why not say....remove the 8 edges of the cube which are less time/space relevant in terms of information in relation to the central pixel? Much more simple...
Can you explain about classification of whether bike or car using Local binary pattern
Enticing visualizations! I can even mute although his explanation is great too.
Excellent explanation. Can you please tell the simple ways to extract some feature by using MATLAB
I doubt you'll get a numbers up to 288 from an 8-bit number :P
+Nathan Prins Not if you only have one set of 8 bits to work with. But if you have much more? You could use some tricks.
Bungis Albondigas I was just saying that an 8-bit integer can only have a value of 0 to 255 :)
***** Yes, that is true. Alone, the counting rule applies to the data, with no compressions possible accept for one to one mapping with a few small values.
rare video where paper is being used instead of cardboard
Could anyone please tell me the difference between image classification and image retrieval? Anyone please...
I want PPT for Change Detection in Optical Satellite Images Based
on Local Binary Similarity Pattern Technique can any one send plz sir
How do you make these animations ?
Thanks very much for your great video!
Really Neat explanation, Thnx :D
Please make a video about self driving cars. There are a lot of conflicting opinions about whether or not we should allow computers to operate a motorized vehicle on public roads, and there has been some news of cars being hacked. As a student of Computer Science, it seems to me that if done correctly it should be safe, all the same I'd like to know the opinions of others as well.
Ok for getting into the thinking of image processing but that technique is nowhere robust enough to handle the real world.
The short answer is neural nets.
The number of blocks doesn't go down from 26 to 8, it goes down from 26 to 18. No big deal though.
+Ethan uhm.. it reduces the search space from 2^26 to 3*2^8, the number of blocks encoded was not the point there.
(just for for those curious, 3*2^8 is 768, 2^26 is 67.1 billion, quite the reduction)
+NeXtDracool Long or short scale? It matters you know...
EDIT: Actually... No, i was too quick in replying. It's ~67 Million either way. 67,108,864
+Ethan it does go down to 8, because it only considers dimensions 2 at a time. Each pair has 2^8 possible transitions, and 3 dimensions can be paired in 3 ways (XY, Xt, Yt), so this orthogonal method has 3*2^8 possible values. Don't be tricked by the image, the method is not just taking away the edges
I didn't get how the search space was initially 2^26? Can you please explain this to me?
I get the reason it's 26 (9+9+8) but why 2 as the base?
Thanks for the video, I have a question. is this local binary pattern invariant to rotation and scaling?
thank for amazing content
It's always difficult to listen to fellow dutch colleague, due to the accent. I always hear the dutch influence. Still a very good explanation though!
I was told there'd be Tom Scott (Inserts meme here)
damn this guy can drone on about nothing.
well explained..
really excellent explanation really got this! thank you!
Is it just me or does he have ridiculously short index fingers compared to the other fingers on his hands
i want matlab code for skin cancer segmentation using thresholding
And I want a million dollars
@@dinnerplanner9381 nike trick
more please
Even though we've achieved amazing progress on object recognition in computing, we're not even coming close to a viable solution about it. I'm truly convinced that mathematics is NOT the answer here, as is not a good answer for every other NP problem.
But, the video is great to show the very reason why - at any point, the algorithm is "aware" of only one single dot of reality. And a pixel is meaningless in the real world.
+OnE61811301 If not mathematics then what? magic?
+Peter Walker I don't know :) If I knew, I'd probably have several Nobel prizes on my chest. Maybe some magic branch of mathematics' paradigms and models we haven't yet dreamed of. What I'm saying that the current mathematical and computing models of "real world awareness" is totally useless. Well, not totally, but far, far away from anything close to real-time.
thanks alot
first
He's Dutch... sorry maar je hoort het direct...
288!! WWHAT?
67108864
second