Comparison between YOLOv10, YOLOv9 and YOLOv8 on Real-world Videos

so in this video here we're going to do a comparison between YOLO V10 YOLO v9 and YOLO V8 so we have all of these models available now all the members in the YOLO family and we're going to do a comparison both performance- wise but also speed-wise so this new YOLO V 10 model is very interesting because it does some optimizations and removes some post processing steps that we're going to take a look at so it is act like significantly faster compared to the other two models but let's goe and see if we act like take a hit on the accuracy so to start with before we jump into into some video examples let's just go inside the YOLO V 10 GitHub repository where we can see these comparisons on the Coco Benchmark data set and if we take a look at it it is clearly the best model YOLO V10 but later on we might go in and see some other things once we see some image and real world examples from wi so we can see first of all here it is significantly faster compared to the other models we have YOLO v9 here in the other red and also YOLO V8 so this is the green one up here so it is act like twice as fast compared to ulv 8 and comparable performance even the largest model here it outperforms it on the Coco data set also number of parameters is significantly lower so that is also pretty cool and even for pretty much all the models here it doesn't really increase that much for the large and also the extra large version where some of the other YOLO models here you can see that the number of parameters here pretty much just increase exponentially with this size of the model where it's a bit more linear here with the new Yol V 10 model so this is pretty interesting let's now go and take a look at what are the differences with the new YOLO V 10 model and if we go inside allytics documentation we have all the YOLO models in the YOLO family all the way from YOLO V3 to YOLO V10 but we also have Sam mobile Sam Fast Sam real time detection Transformer and so on so you can go and read about all the individual models in here see how you can use them take them with just a few lines of code and use them in your own applications and projects so you be 10 here the main difference is that they remove the non- maximal suppression in the post processing and what that means is basically like when we have our optic detection model it act like outputs a ton of different bounding boxes at the output layer which is basically just guesses of where the bounding boxes are so we can probably have like five six seven eight 10 bounding boxes on top of each other having different overlap and then we go in and do non- maximum suppression which basically just means that for all the overlapping bounding boxes with some threshold with intersection over Union we're just going to remove those wounding boxes and merge them together so we just have a single one covering our object so this is the step that they're removing for this new Yul 10 model and that is act like where it gets a lot of the speat UPS from compared to the other ones but also some other optimizations in the layers so they also have some design optimizations in for example the classification head which is a bit more light weight compared to the other YOLO models but also when we're basically just talking about like the layers they have some down sampling layers and so on here and there which is a bit different compared to the other models but again removing this non- maximum suppression in the post processing reduces the processing speed significantly because this is act like a very expensive operation especially when we're talking about like milliseconds and often if you export the models and so on make sure that you act like include the non- maximum suppression in the model and also does optimization on that part because if you're just optimizing the model and then you need to do non- Maxx impression on your own you won't even be able to run in real time if you're doing this non- maximum suppression in Python so those are some of the differences for this new model or at least the most significant ones let's now go and see some results on videos for all three models compare them and talk about the results so I just have these videos here where we have the y v 10 model on the left side and the y v 8 small model on the right side let's go in and see if we can see any differences on these real world data sets so this is basically just videos running so this could be real world and real life scenarios and then also have the video for YOLO v9 just to see if we can do some comparisons and see if there's any differences so let's just play it here and let's take a look at the results so right now just initially it looks very similar I can probably see that is a bit more dense with the prediction just on a high level for the Yoli 8 small model compared to the YOLO 10 model on the left side so here we can see that we're detecting a person sometimes we're also detecting like a traffic light down here for the UL V 8 model where we're not doing that with the ulv 10 model so right now just initially by watching the video through a couple of times so it looks like the Yol V 10 model has less false predictions so again we don't have any persons here we don't have any traffic lights even though this is a traffic design it shouldn't be labeled as a traffic light so right now let's just play it again we can see directly that Yol V8 act like does a better job on smaller objects compared to the Yol 10 model and this could be because they're reducing the number of param and they're also having a more lightweight classification head and so on so this is my experiences with yolo V 10 it is not as good on smaller objects compared to Yol V 8 but again there's always a trade-off between accuracy and speed when we're lowering the parameters when we make the models faster we will always lose some accuracy and so on at least for the yal models when we're comparing it right now so these are my general experiences like we can see we'll miss a lot of cars over here to the left and also some of them driving to the right here on the highway we basically just detect all of them up here at the top right corner with the UL 8 model and we do the exact same thing with the Y 9 model which we're going to see in just a second so Yol V 8 better on smaller objects Yol V 10 is significantly faster compared to Yol V 8 it is up to 1 and a half to two times faster maybe like 50 100% faster so it really depends on your applications and projects if you really need to detect small objects or you prefer to have a faster model if you take a look at it down at the bottom again we can see of course we can go in and lower the threshold more for the Yol 10 model we will probably get more detections so it looks like initially that for Yol 10 we need a lower confidence score compared to Yol 8 to get more predictions and that will probably also introduce some false predictions here and there but these are actually like running with the exact same confidence score just to make the comparisons as even as possible down the bottom video we can see that we detect a person here again that is not correct we detect an umbrella with okay high confidence score we even have some bosses up here at the top which is obviously cars that is pretty easy to see but the problem is we detected as boss with yolo V 8 but with yolo 10 we don't even detect these as cars so these should be pretty clearly cars again we can probably lower the confidence score but the confidence score is 25 so that is still very low of course these model will be significantly better if they fine tune on a specific class so if you take all the cars label them and train the G V10 model fine tune it on specifically cars it will be able to detect all these cars no problem these are the pre-trained models on the Coco data set and we have 80 classes so let's just run it here and let's see if we see some other differences right now we also got the pur detection down with yolo V 10 and we also have some other wrong ones here so these are trucks so label seven is trucks we have bosses here on the right side again now we can see we get some trucks on the right side with yolo V8 as well we have the persons here it's figering back and forth it thinks that this is a person person but it also does that now for YOLO V 10 sometimes so yeah again these models are pretty similar you can really just take one of them we can just say that okay now we have a new Yol V 10 model this is the best model it shows the best performance on The Benchmark and because of that I'm just going to use that in my application and project it really depends on your data set the problem that you're trying to solve if you wanted to take smaller objects if you're running this on Lower End Hardware so you really need the speed if you can get away with just like 10 20 frames per seconds probably even more depending on the hardware you can get away with u 8 and then have a bit more accuracy of course you can always go up to the medium model compare the medium model to the small model for y 8 and you probably be able to get more detections but again you can see here it pretty much just misses these cars it shouldn't do that and the confidence scores are also generally lower but this comparison here is pretty much just even confidence score every single parameter even the models here are the exact same ones but again I'd probably rather go with the medium model for YOLO V 10 if I were to use it over a YOLO V 8 model if we take a look at the YOLO 9 model just to see the results from that one it does a pretty good job with detecting smaller objects here in the top left corner and we also get some false predictions here and there but that is kind of expected the Yol 9 model I feel like it's a bit better on smaller objects also on all the other experiences videos and all the models that I've trained on my own it looks better when we're act like doing this on smaller objects so this is pretty cool definitely go Ahad and test it out in your own applications and projects I'm just going to let it run here for the video so we have some false predictions with traffic light person but generally it detects pretty much all the cars we can even see the cars at the top coming into the frame significantly better compared to the Yol V8 model and again this is a really hard s case this is probably like the the hardest that I can come up with for a real world scenario but it shows pretty good comparisons between all the different models and it is the exact same experiences that I'm showing now that I've had with all the other models even custom models and so on that I've tested out if you want to know how you can use these models definitely check out my channel I have tons of videos covering all of them even how we can apply optic tracking on top of these optic detection models how we can use it with Alo litics train our own generate our custom data set run inference in a custom python script even export them and do optimizations so thanks a lot for watching this video here guys I hope this has given you some insights into which model are the best right now and again again there's no answer to that it really depends on your application and project Yol V 10 is not the best model just because it is the newest version make sure that you test it out on your own and then I'll see you guys in one of the upcoming videos until then Happy learning so we also have an AI career program if you want to learn how to land AI jobs and get AI freelance work I teach you everything in there we have programs all my technical courses weekly live calls personal help and I would love to have you guys in there help you out in any possible way you can check out the program down description and the community and then I'll just see you guys in there

Channel: Nicolai Nielsen

Views: 2,894

Keywords: yolov9 custom obejct detection model, yolov10, yolov8, yolo version comparison, yolo versions explained, yolo comparison, yolov10 vs yolov9, yolov10 vs yolov8, yolov8 vs, yolov8 vs yolo v9, yolov10 vs yolov8 vs yolov9, best yolo model, how to choose which yolo model, comparison of yolov8 and yolov10, Yolo models object detection, Yolov10 the best model, best object detection model, yolov5 vs yolov8, yolo v10, object detection, yolo 10, yolov10 object detection

Id: x20MxX-AWzE

Length: 10min 34sec (634 seconds)

Published: Sun Jun 09 2024