New YOLOv8.1 Object Detection Model with Oriented Bounding Boxes

so Yol 8.1 has just been released now we can go and do RN bound box detection with allytics and YOLO V 8 so this is a brand new model we're going to take a look at how we can run it in this wood here we're going to go over the documentation see talk about like what are R in the bounding boxes why there's a big breakthrough in optic detection what it can be used for and then we're going to see a couple of examples so let's just jump straight into it and get started so we're just jump straight into Al latest documentation here we can see that we have these Arn Bounty boxes with the new Yol 8.1 model because right now Y8 turns one it has been out for uh one year I have multiple different videos here on the channel definitely go check that out see how you can train your own custom Y8 models you now you can also go in and train your own optic detection models with oriented Bounty boxes but you can use it for segmentation tracking post estimation classification optic detection and so on I have videos about all of it so why is this breakthrough in optic detection and the reason for that is that now we can actually like fit our bounding boxes better around the Optics that we're trying to do detections of so right right now if I just zoom into this image we can see that this is an image from a data set so we have these boats here that we want to do update t on let's say that we're just using traditional bounding boxes then we'll just draw like a rectangle around the bounding box it can really be rotated so that will just be a fixed rectangle or square that we're drawing around our uptic but now we can actually like rotate it or Orient it around our uptic so this is a big breakthrough in optic section and now we can use it out the box with ultr litics the hold framework around Yol V8 and so on these models here are easy to train and also running samples on so this is pretty cool we're going to see how we can run inference and then we're going to see some results later on but just go or documentation here this is the data set that have trained the new models on so all the models they have pre-trained models available you don't have to do anything it's only just a couple of lines of code that you need to write and you have everything up and running as with the previous y V8 model they also just added this new cool feature the new ultral litics explore where you can basically just go in and explore your computer vision data sets you can do semanic search SQL queries similarity search and so on I also have videos about that here on the channel where we're using tinix but now we can go in and search and basically like try to like f figure out how your data set works you can go and search for for example let's say you want to search for taxes or like cars in your data set if you want to fine-tune your computer vision data sets because that is very important when we're working with real life applications and also projects so again the models are not too important the data is significantly more important when we're creating models nowadays and also basically just putting them into production so better data is better than better models but yeah this is also a pretty cool feature we're going to cover this in another video in this video here we're just going to cover Yol 8.1 for oriented bounding boxes so if we go inside the task here we can go inside obb for oriented bounding box object detection and here we can basically just see the exact same documentation as they have for the detection segmentation classification and also post estimation so these are all different tasks that you can do with Al litics and we have videos about all of that here on the Channel they have five different models again they have a nanom model small model medium model large model and ex large model where we're basically just changing the model size so in the M number of parameters as you can see over here you can see some different metrics both for CPU and also GPU mean average position test so these are some pretty good results if we go further down we can see how we can train it these are like single liners that we can use Alo litics for so import Alo litics or from Alo we import Yello create an instance of our YOLO model and then we can just directly call train predict or what whatever you want to do whether we want to train our model on a custom data set fine tun it we can also train it from scratch or if you just want to do inference with a pre-trained model we can specify a number of different parameters we have covered that in all the previous videos but here we can basically just see how we can do a prediction and copy past the code here throw it into python script and just run it you can throw in different images videos URLs and so on so we're going to take a look at that in just a second you can also export it to different formats like on next torch script pytorch open B T RT and so on if you're going to deploy these into production so before we're going to jump into the code I've just found this video on YouTube we can copy the URL paste it directly into the predict method and then we can do inference on this YouTube video directly you can throw in like an nonp array video specifi The Path video video stream webcam or like whatever we have covered that in the previous videos so now we're jumping into the code so we only need to set up these color lines of code so from allytics import Yello then we can create an instance of our YOLO model where we just specify y V eight large you can also use medium small or whatever we're just going to go with large in this example and then we need to specify obb for oriented pounding box. PT if you just want to do traditional update detection you just delete this part here you can also do segmentation so that will be SEC and so on so this is how easy it is to use with ultral litics so now we have created an instance of our model we can now go down and create and do a forward pass do a prediction with our model that we create instance of so this is how we can do it we can also call the predict function so there's number of different argument you can specify it into section of Union confidence scores if you want to show the results and also if you want to store the results the device that you want to run the inference on and so on so right now we're just going to use the CPU we're going to show it we're going to set the equal to true and then we can throw in pretty much like whatever format that we have our data in could be video URL um nonp PL image and so on right now we just have the YouTube video let's run the program here and see how it works so first of all here I'm just going to activate my cond environment after we've done that let's go in and P install allytics if you already have it installed make sure that you upgraded for getting these new um functionalities so we need inip install ultral litics and there we go we can see that my requirements are already satisfied but once you have done that you have everything so this is three lines of code one pip install and then you have RN bounding box optic detction up and running so now just go down and run this so Python ob. py and now we should be able to get the receed results directly and we specified show equal to true it will download the model automatically as you can see here if you haven't run this program before after that it will be stored over here in your directory so just in a second here we should be able to see our window 85.4 million parameters now it's going to open up the Wier stream we can just see that sometimes we're waiting for a stream because we're pulling this from the YouTube video but we can see a bunch of different things like how we can extract the result so this is just a generator that it's returning you can extract the boxes Mass probabilities classes confidence scores and all of those things but right now here we can basically see we're running this large model and this is running on CPU so let's try to go down and see if you can see the inference speed so that is around like 250 milliseconds so that's around four frames per seconds that we're extracting now let's take a look at the results here so this is like some Drone footage of a traffic intersection so sometimes we actually get predictions of um a ship with act like a pretty high confidence score um even though this is act like a car but it is also mainly trained on like Harbor data so like ships containers and and those different kind things for the pre-ra models again you can go in and fine-tune your models I'm going to create additional videos for that so how can you take your own data set how can you label your data set in oriented bounding boxes and then how can you train your costom model here we can see already in the bounding box when the car here is taking a right turn so before we will just have the traditional bounding box the static square or rectangle so now we can see that it is turning and the bounding box is actually like oriented or rotating together with the Optics so now we have bounding boxes now we have optic detection that is fitting way better to our data so this is pretty cool I'm really excited for this I'm going to play around with this a lot and we are going to create a bunch of videos here on the channel so definitely make sure to hit the Subscribe button on the video this is a breakthrough in optic station it can be used for a lot of different applications and projects and this is way better compared to just using the traditional update detection it will basically improve every single update detection project out there like there's no reason for not using these oriented bounding boxes and you saw how easy it is to run with ultra litics this is the new Yol 8.1 model just released this is pretty cool we're going to play around with it a lot try it out yourself try to see if you can come up with some cool projects I hope to see you guys in one of the upcoming videos until then Happy learning

Channel: Nicolai Nielsen

Views: 7,370

Keywords: yolov8, yolov8 neural network, yolov8 custom object detection, yolov8 object detection, yolov8 tutorial, object detection yolo, object detection python, opencv yolov8, opencv python yolov8, object detector, object detection yolov8, opencv, detect objects with yolov8, yolov8 opencv, opencv dnn, deploy yolov8 model, obb, object orientated bounding box, yolov8.1, yolov8 obb, object orianted

Id: s4YesBn9bPM

Length: 8min 17sec (497 seconds)

Published: Thu Jan 11 2024