YOLOv8 - Multiple Object Tracking (MOT) | Supervisely tools | Computer Vision Tutorial

hello my name is Max I'm CTO and data scientist at supervisly and in today's tutorial I will show you how to perform multiple object tracking in supervisely on your videos let's start this is my supervisory account and I will show you right now how to perform multiple object tracking using the YOLO V8 model so here's my data as you can see there are a lot of frames and a lot of cars on the video stream and we will try to detect and track all of them let's go back to the supervisory dashboard to run neural network unsupervisely first of all I need to connect my computer with GPU you can do it here in team cluster as you can see I have several computers connected I will provide the link in the description to this video how to do how to connect your own computer to to your account so let's do the following let's go to the neural networks page find the category object detection on images uh section serve and deploy Elevate model on my computer I need to press run a button go to advanced settings select computer I'm going to use for deployment and press run right now the application is started here we can provide our custom checkpoint but for this tutorial I will use already pre-trained model that is available in community I will you I will pick the large model object detection and will deploy it on this device let's press serve button so right now the model is downloaded from the GitHub and deployed on my computer let's go back to our demo data to apply detection model to our videos we will use special app from supervisory ecosystem it's called apply neural network to videos you can find it in neural network section here let's run it right now the application is ready and we can open it these apps do the following it uses a paradigm tracking by detection here is an example we have a video sequence our detection model will be applied to every frame separately and as a result we'll have the separate detection on every frame then we will use some tracking algorithm in our case it will be a deep sort automatically applied and thus we will combine all this separate detections into the tracks and final trajectory and get the final trajectories for all our objects let's go back to the app here we have our images let's press select button and on the next step we need to connect to one of the deployed models in my case I deployed a few minutes ago yellow V8 and I will connect to it we see some basic information about connected model on the next step we see some basic classes that model can predict let's select all of them and go to the next step here I can configure inference settings like confidence threshold for all prediction intersectional reunion some settings for non-maximum suppression algorithms and so on let's keep all the settings by default and press preview button to see the tracking result on the short video fragment so as you can see the preview is generated and we can watch it before we apply the model to all videos in our project let's define the name of the resulting project let's call it inference results and press apply settings button on the last step we will press this button to apply the model to all videos in our project in my case for them reasons I have only one video but you can have hundreds or even thousands of them let's wait until the model will download the video will be applied to every frame then the tracking algorithm will automatically group all separate predictions and build a tracks for our objects let's wait until the app will be finished the model downloads the entire video and then iterate over the frames and predict objects on these frames and then at the end all this predictions thousands of objects will be combined and uploaded to the platform so you will be able to preview these annotations on top of your videos in supervisory video labeling tool as you can see the model generated about 60 thousands of objects on our video and it will be really hard to label all these objects manually thus these applications will help us to automatically pre-label the data and find objects and set tracks on our videos so right now the application is finished and as a result with a new project with inference predictions is created let's check the results so that's it for today I recommend you to try to upload your videos and perform per labeling quiz the portraying or custom model in supervisory if you find this video useful subscribe to our channel to watch more computer vision tutorials if you have some questions please leave them in the comments to this video see you have a nice day goodbye foreign

Channel: Supervisely

Views: 1,027

Keywords: computer vision, tutorial, how to, neural network, gpt, gpt-5, chat gpt, openai, ai, artificial intelligence, nn, workflow, annotate, label, tag, industrial data, track, object, video labeling, annotations, multi-object tracking

Id: t6qsFMjB_xs

Length: 7min 27sec (447 seconds)

Published: Tue Aug 08 2023