 Good evening, everyone. We are here to present the SyLab computer vision toolbox. These are the name of the team members I'll quickly skip through this part So our aim our aim is to build a comprehensive computer vision toolbox which is at part to its software counterpart Which is MATLAB. So what we've done in this toolbox is we've included various Computer vision algorithms. We've tried to implement feature detection extraction and matching Object detection and tracking, motion estimation and video analysis and camera calibration and stereo vision A brief introduction to SyLab. Okay, so these are the names of the few open-cv modules that we've used First one is core. It includes the basic data structures like MAT which is the container for an image It can also contain some basic image processing algorithms and then Higuie which contains the graphical user interface which openscv uses IMGPROC video and other such modules So we've also used Tesseract. We've implemented optical character recognition in our toolbox We'll just start with the demonstrations. The first one is the image category classifier This is the image set we've used. There are three categories here One is of airplanes, one is of ferry and other is laptop. These have been taken from the California Tech Technological University's standard computer vision data set The first thing there is the image set. We are taking image sets. That's the image set which I just showed you. We formed that into a Image set structure which can then be used by SyLab. We are splitting it into two parts which is the train set and the test set It's a 0.8 split which means that 80% of the images will go to the training set and the remaining 20% of the test set We've created a bag of features visual model which helps us to categorize the images into which category they belong After that I'm training a category classifier on the basis of the training set which I generated It's been trained for three categories which are the categories there in the image set The next thing to do is evaluate this on the test set So it's been evaluated on 13, 16 and 13 images from the three categories respectively And we're getting a very high accuracy But this is again because it's a very small data set and to include a huge data set We might have to include more complex image processing algorithms How many features you have extracted for the training? So we've extracted quite a lot of features. It's over I think 97,000 But we've clustered them into 500 centers and so basically the output is generated on the basis of just those 500 k-means clusters LCA I'll give you an LCA image. You know what is LCA? No sir No? You've worked so hard for the past 33 years in such a big India And you haven't taken that image for the test? Will Modi do it? And hopefully you will classify it as an airplane I don't think you will. That's why I asked It doesn't look like this Okay, sir, I'll try What about the stealth bomber? Sir, I have not included those I have not included those in the image set You have not tested it, no? As I said that image set is of the California tech Yeah, but the stealth bomber image is not classified on the net Sir? It was printed on the LCA paper What's the set for that? You can add images to the site Okay, so I'm just showing you how this prediction is working I've read the image of an airplane and the label out there we can see it's detected as an airplane The next example, this is a laptop and it's also detected successfully as a laptop And then this is a ferry, I've just shown three examples Okay, now I'm going to continue on these image categorization examples And demonstrate recommender systems These systems are used in e-commerce websites like Amazon and Flipkart And they are used to recommend similar products so that user can buy them So this is the dataset that I'm going to use These are the different images of clothes, wedding gowns, t-shirts and all those suits So I'm going to train these images and recommend similar products So first it is creating an image set consisting of different clothes So all the things you showed me are training set? Yeah, those are all the images that I'm going to train on And then we have created a bag of features This is a bag of vocabulary words that we are associating with each image And now we are creating an index of images in which each image that is present in the dataset We are associating what features in the clusters are present and have similar properties So what we are doing is feature matching Now this is the first image that we are going to test This is the first query image It is an image of a t-shirt So this is the first image in the entire dataset So it is a t-shirt But why not light combat aircraft? We can use that It is not a play, can it tell me a problem? Yeah, it can be But this is the The idea is not light combat aircraft The idea is you have to do some destructive testing on your algorithm Okay That it recognizes a t-shirt okay, it does But does it recognize something? It is a query image, it is not a recognized image This is the image that we are going to search What are the similar images? Correct, I understand If you are showing me a t-shirt and saying my algorithm is successful Because I have extracted features and matched Okay, that is not sufficient for your algorithm You have to prove that if I give a camera It does not classify as a clock Sir, it is not a classification system It is a retrieval system What you are going to do? You are retrieving similar images If in the dataset there is one t-shirt picture And there are different airplanes So it will, you know, extract those What is the purpose of this feature machine? Sir, recommender system If you software it, call it an application Applications? Sir, entire industries, Amazon, Flipkart All use these recommenders system This is an Amazon and Flipkart algorithm Which when I say close is not going to give me a camera Yeah So your algorithm might Yeah, no, it not, it will not That is not tested against the camera Yeah, it is tested sir It will give you a return How it is tested? You are not, show me Have you tested it against a non-cloth image I am not stating no This is the t-shirt image It will not give you wedding gowns or, you know, suits It will give you t-shirts only I am not talking about t-shirts only Okay I am saying I give you a random image Does it classify it as an evening gown? No If it does, Amazon doesn't, Flipkart doesn't I know Do you? Do you not test it? Where is your destructive testing? This looks like that t-shirt only How do I know that you are not classifying it as that? That is what I am trying to say You, again it comes down to the single fact Okay You people do not do testing Okay I will not allow anybody to claim that my thing is okay You are even the discordant, I classify it as t-shirt I will actually take this Where is that kind of testing? I don't see Sir, this is not a classification algorithm This is a retrieval algorithm Correct So retrieve, if I give you this And you retrieve a t-shirt, is it good? No So you will Sir Is what I am saying You are not tested If I may intervene here If you give me the photo of a phone And if I have images of phones in my data set It will give you the images of phones I am another t-shirt You are all one Sir, yes It will match as a t-shirt You are not tested I am not saying it will Where is your destructive testing? Sir, if you are using the entire set of If you want to recommend mobile phones t-shirt My question is not Don't give me this Don't give me LCA Where is your destructive testing? Sir, even in Amazon There is no alien photos If you will give an alien photo You are not using software developed by interns Okay Don't talk to me about Amazon Okay You are Again you are saying I am developing software We are equal to that done by Amazon You can I am not saying you cannot But you need that level of accuracy Where is that level of accuracy? You can't write algorithm without testing I will not accept Sir, these are Sir, this is already classified in K Sir, if you want This is a new image This is a new image He is going to look at this image And from his closed library He is going to find out t-shirts And he is not going to find out evening gown That is his success And that he has tested What he has not tested is If I give him a phone Whether he takes evening gowns Or t-shirts Sir, if you That is not tested So, he has tested With a set of clothes And he has tested that With a set of cloth images What I am saying is Any testing has to have destructive component Non-cloth images have to be tested Sir, if Which look similar I am not saying tested with LCA LCA has got this kind of shape That is why you may pick an evening gown Once again, that is the classification algorithm Whether it is that t-shirt image Forget all this t-shirt It's a business Tell me, you have got a set of 100 images Yes, sir I give you one image You are doing all kinds of You are doing all kinds of your 1950s Whatever pictures you have What is the output of your Sir, we will give you the image Which is closest The similar, most closest match The closest match to your phone We will give you that image as an output That is what our algorithm does So, for this you will give me a t-shirt Sir, if this matches a t-shirt Then we will give you a t-shirt Flip card doesn't do that So, then don't mention flip card in Amazon Okay, it gives me a set It does not give me closest match First of all Sir, we are also returning a set of closest matches Correct So, take your words back I will give you a set of closest matches Not one closest match Okay, yes, sir We apologize That was a mistake Correct We are actually demonstrating We are returning three closest matches Three closest matches So, suppose I give you this Okay I give you this That is your function Yes, sir Have you tested that it does not return me Three closest matches as evening gowns How do I know? When there is no destructive testing I cannot know You have to get it into your head You have to do destructive testing Sir, one Half my time is done in destructive testing Only I have told people Who work with me That if you tell me this software works Do not expect me to check Verify that your statement is correct I don't want to verify I will only verify how your software behaves When I do non-standard things That is the quality of software Nothing else is quality of software Okay I have never Those who have seen me test I have never fed the correct data Never in my life I have fed the correct data to any piece of software That is given to me I am assuming it works correctly If I give t-shirt it will pick up t-shirts I assume That you are not idiots But I don't know whether you are smart Smart will be there There is no t-shirt I don't have a closest match Do you do that That is the question Sir, the thing is There is In this concept We do not have a Option of not a closest match We will always try Even if it's a 1% match And then a 0.5 And then a 0.25 We will return those 3 That these are the closest 3 matches What we can do is We can set a threshold That if it matches above a particular level Then only return Otherwise say this is not a connection Why are you having close connection? I really don't understand You could have put airplanes also there See, I didn't understand Classification and matching In principle there is only Getting some information about the Image So why you are unnecessary Putting it separately Once I understand Whether it is a t-shirt I may classify it Or recommend it Or whatever So all and all Ultimately it is only one same thing If I have to If I have to purchase an airplane You will show me the airplane over here Correct There also for classification You are saying the airplane also So why unnecessarily You are repeating the things Ultimately technology is one The thing is This percentage of matching The classification one was If I give you an image And I ask you What are the things present in this image That is done by a classification I will extract sections of the images I will compare those images in my data set And maybe let's say There are a few dogs And a few cats in one image So I tell you These dogs are there And this many cats are there Ultimately behind the screen It is the same thing Only it is the only application That either say it is being classified Or it is being recommended Different applications of the same So why unnecessarily spend the time In showing the application Until the technology Go ahead Right Okay I will just give them one example Face recognition people will be unhappy If I don't do that Okay The first time they came With that face recognition application Okay I went and borrowed a magazine Okay They expected me to check Okay In that magazine I gave various images Okay So far it was Okay I expected it to work In this case Whether it picks up faces from the magazine Okay Then I gave them a photograph Of what Smuti Irani was What was Narendra Modi Smuti Irani and Narendra Modi Had a stage photo 10, 15 people With all the things I fed him that What does software do? You know what they detected Flower pot That is what I keep Screaming at them for one week That you are detected a flower pot Okay Do they use software? Software said This is a flower pot There are two ears There is nothing else Okay So the flower pot Made with two ears Made with a whole face This square I can see That is what I mean by testing Okay Actually happened So I am continuing with this So this is the similar images Then this is the test case Then it will fetch similar images once again And this is the other test case And once again If you remove the tie Will you see the coat? Hey If you remove the tie Will you see the coat? Yeah it will give So it will give the most similar match If you remove the tie Will you see the coat? Yeah it will give I have not tested Test So I have done testing But not the closest match If the guy is wearing a tie Then the next Closest match in the evening Sir closest match will be The person who is wearing a tie With the coat Yeah it will not be a closest match The person will Is that testing? Yes sir Is that testing? Sir that is tested Because it is finding the most similar image So if you remove the tie Will you see the coat? Yes sir it gives I have tried to implement foreground detector Which is the primary step In motion detection Especially in applications like Traffic surveillance system So I would like to just pay the video Right now Yeah test 3 Just separate it This is a flame Initially there is a ball In the top right corner But when the algorithm realizes That it is no more a part of the background That goes away So this is what the algorithm does I have run it on many videos And well right now The So it detects the foreground It sets a background model With the initial number of frames And then the It removes the background And it is done using E.M. algorithm Definition of foreground Pardon? Definition of foreground Something which moves in a video If there is a video All things which are moving Yes sir So this is the primary step You are moving on to the next term Excuse me What features? So Well This foreground detector Is implemented using Gaussian mixture models We have many models Gaussian models Like For example in a traffic surveillance system There will be a Gaussian for road What will happen if I move the camera? If you move the camera If you move the camera Then we will have to start the foreground detector Again Because it has to set a Background model So that For each individual pixel which comes in It has to subtract it From the static background model So we have to Set the background model first If a person is walking Yes sir And I am following him You will not detect the movement Why can't we We will If the foreground detector has Moving camera We will not be able to reset it No sir If the background model is set And it comes in the frame It will detect Background model You start walking I am taking your video Walk here I have taken your video You have created it Will it show you or not? It will definitely show you So background is moving No the background You detect whether the movement Of the image In any change Is because of camera motion Or because of the No sir this is specially For traffic surveillance system Where there is a static camera Where there is I do not ever agree That there is anything special I understand there are Only traffic camera Foreground detection system Okay All right If you are talking foreground And background The concept is If I move the camera How do we detect foreground And background How do we differentiate The point is It depends on the algorithm It has to be given some time By the way To set the background I am going to question my video team Okay And they have used One of the Your own libraries Okay Which defines a background On the fact That a camera motion Okay And a background Which is stationary Okay There is some Function Yes sir Which color Or all the image Saying How many Images have moved In what direction Right Majority of the images Okay So Background Say you are talking Say you are talking about optical flow Where That is used for Okay That is That according to me The definition of background Yes sir Another definition of background Okay Sir these are That background Is something Which moves the same way In the camera If it is a majority Of the picture So the point is What are you talking about I can identify So the point is What are you talking about Optical flow Which we have implemented Inside this If the camera moves Inside this Not right now It has to be integrated With this Sir these are different functions Here You have used the Frame differencing Correct Frame differencing And then thresholding Yes sir Yeah So it's okay It's standard practice Not a problem Okay Good Thanks Sir next Demonstration is on People detector in Psylab So this will detect People in an image And it will Give It will be as close as To people So this is one of the Image in which It will detect Three people And Next images So we are using The library function From OpenCV Okay It already detects What are you talking about Sir it is implemented In Sylab It is What is the function In Sylab With OpenCV command And you call the function Is that Yeah I have seen this Hey Other people What are you talking about Back back back Sir I will Now we Will show you How to Get it Sir I will I will not do it Right now We have tried to implement As many functions as we can And we will be Implementing on Making these functions Better Right now we are Directly using the Model which OpenCV gives us And we will be Implementing it And improving it further The thing is that You people have Only implemented The OpenCV library You have coded something Because the Other batch I was very impressed This guys has come up Only three functions But their own coding So here We have used Direct OpenCV library Implementation Or your own coding So I will be very happy If you do Your own coding Other than the Pulling the Readimate library There is a Algorithm called Algorithm Channel Features Which is not present In OpenCV Okay So we have implemented That also Okay So that Okay Right now MATLAB Does not have Ownership of In Software It is always Person Weeks Don't say Three weeks If it is Three persons Then it is Three persons week He has done Okay One person Three persons Yes sir So okay Good