Image Labeling using Active Learning reduces Human Effort

hello and welcome to another video brought to you by the bridex ai accelerator team at samsung sts my name is patrick bungard i'm the vp for artificial intelligence and today i'd like to talk to you about labeling image data you might want to classify an image into categories draw a rectangular bounding box or draw a free form boundary around a segment of an image when you label an image data set like this you typically label it in whatever order the data set came in now if you were to label the images in the order of the information gain that they provided to the artificial intelligence model you're trying to obtain then you would be able to save yourself a lot of effort the way this is done is that you first label a small amount of your data and provide that to an ai system called active learning which then tries to label the rest of your data automatically it provides you feedback in the form that it gives you another small segment of data to label namely those data that the system is most confused about you label them and provide it back and again the system provides you the data that it's most confused about and after several of these back and forth iterations the system is no longer confused and can label the rest of your image data automatically we experience in practice that you can reduce the total amount of human effort by between 80 and 90 percent samsung sds has teamed up with the startup red brick ai to bring you a beautiful user interface as a sas service with state-of-the-art mathematics to do this process right let's have a look at this in action so here we see example of our data set which is just cars in multiple settings we're going to want to identify those we'll create a project here which is going to be of the type auto label which has the sequence that you see here we start with an input there's an active learning step it gets labeled and there's a feedback loop until we're happy and then output we're going to call it auto label demo and create the project by first of all telling it which data set we're going to use what taxonomy it's going to be and what label set we are going to want to use there's our project we open it up and it's being loaded now we see that there are 100 000 images pre-loaded none of which are labeled at the moment it says zero labeled images and we can start this process by so-called sending a batch which means that now we obtain a batch of images to label it's going to be 1 000 images to label at each time and here we open up the queue of those first 1000 images that we have to label there's our first image and we start labeling it it's a car and that's the bounding box around it we submit that task we go to the second image and we can see that there's a nice trigger here to allow us to label nicely again we label that bounding box to be a car and submit the image and we keep going throughout the various images that we have by labeling these cars among the list of images you can pick the one you want to label next if you don't want to go in the queue that's being prescribed here we just go through 1000 images until we're able to start the active learning process as you can see it's quite convenient and quick to perform the actual labeling and after we've labeled 1000 of these images we can trigger one cycle of the active learning step by clicking trigger cycle this of course can also be automated at this point we can see which labelers did how much work and and see who's to be attributed after the first batch has come back we can see what accuracy has been achieved by the models it's not great yet but we're only a thousand images into the job so with this in mind we can look and see the confidence that the algorithm has for each of the images and after we repeat this again and again and again after batches of a thousand each time we can see that there is a rapid increase in accuracy from batch four to batch 8 and at about batch 16 we hit the maximum achievable accuracy of 99.8 percent and at that point the active learning process is done the rest of the images are auto auto-labeled and you've just saved yourself a lot of work thank you very much for your attention in watching this video if you would like to try this out on your own data set please contact us at the contact details below and we will set you up with a free trial of this process if you found this video informative and you would like to learn more about the other things we have on offer please give it a thumbs up and subscribe to our youtube channel below again thank you very much for watching another video brought to you by the bridexa accelerator team at samsung sds and see you next time

Channel: Samsung SDS AI Team

Views: 1,854

Keywords:

Id: wcP1fRPKXSU

Length: 5min 49sec (349 seconds)

Published: Tue Mar 02 2021