YOLO v5 Object detection on GPU + Custom Dataset training + Webcam

okay hello everyone hopefully uh we are all good to go uh let me see if it's working properly [Music] perfect so hello everyone today we are going to take a look at the uh yolo v5 object detection on gpu because i recently got a computer with rtx 3090 what is actually the best one you can you can get and i'm so excited to uh that was like first thing i have tried uh i just installed the yolo well originally i wanted to try yellow e3 but then i was like okay um that's kind of complicated and because there was like a lot of things not working for me but maybe that's just me so i decided to check the new one although the last official one was yellow v3 then i think there was yellow before and then there is yellow e5 for this running on the pi torch from different developers but these are working really really well and that's what we are going to do today so i'm going to okay it wasn't good so this i'm going to minimize myself like that and turn on okay almost getting there like that perfect so now you can see me and now you can see my screen uh i'm going straight away into it uh because i have only one hour but we should be good and that's it so first thing uh what we can do is we can oh we can search for yellow v5 of course and in here it takes to github where everything is uh describe how to start and everything i already did that but i'm going to remove uh this i'm going to remove all what is it oh here yellow v5 i'm going to remove that or at least i'm going to move yellow if i to yellow b5 old because i'm not sure if there is something important for me or not to so instead of just deleting it i'm just going to rename it and we can try to follow these steps so first thing we need to clone the yolo e5 uh references jesus repository and uh it's important just get if you don't have git you can always install it thanks to sudo apt install git but as you see i already have it so that's fine then we move to that directory and we can see there is this requirements text file if we show it up there is basically a bunch of stuff would need to be installed like mad pro matplot jesus met protlip numpy opencv blah blah blah but you don't need to do that manually because all you can do is just pip install minus r requirements and it will install everything as you see everything have already installed so obviously there is not much we can do at this point um great so that's it this is the installation but obviously there's not enough for us we want to try it out and we can do that uh by running uh let's try this python uh i was doing this last week so i didn't remember exactly but let's try the source uh well i'm not sure if this will work with my camera on if not i will turn my camera off okay let me make some modification on the because i get some apply okay because i get some warning about the stream betrayed but um i think it should be better now hopefully you can see me well and we'll see so uh we was here uh if i run this obviously python 3 and probably yeah so this automatically will download the weights and fail to open because i'm streaming with this camera so i'm going to turn off my camera on the obs like that and maybe now maybe now well this is something that i wasn't thinking about before so let me remove it and obviously this may be an issue because if i'm okay okay maybe don't say oh here we go so as you can see uh this is the test straight away you can see if it's working or not and that's what i really like um it's kind of fine i mean you see the uh frames per second are just crazy because on this graphics graphics card it's it's just running really well and let's try some objects what was trying maybe phone right yeah cell phone here we are this is really good uh obviously it's not uh you see on the background the monitor is saying it's tv and and this 3d printer uh it recognizes tv but they kind of understand why because you have this rectangle shape with the black bezels that's exit acceptable also the monitor is saying is tv but i mean that's fine this is more or less just to just to see uh how this thing works and uh i think it's pretty awesome for for like first five minutes uh we can kill the script and what i'm going to do now basically we are interested in the training uh because there is a lot of videos on youtube like how you can train object detection and everything so they run it in collab google collab but what is actually pretty fine but then they showed like one picture and then yeah it's trained and that's it but they don't show you like how you can really use it with that game or something usually so that's why i'm here and we are going to do exactly that so i'm going from the beginning but i don't want to bore you with taking pictures or anything so for example as you can see in here i took some pictures with the keys uh i mean for today i'm just going with uh one object and that's that's the keys um i mean pretty boring but it will it will work good this demonstration and obviously i just take okay how many hundred pictures i believe yeah 100 items so that's that should be more than enough and obviously i don't want to bore you with drawing all these uh rectangles around them so i already did that but i'm going to show you how to do it anyway um and you can use uh any any uh labeling i think label image is called or something no this you can use for example this software uh for labeling it's the like the the most basic one and it's working just fine or you have on if you have mac you can use resize me or something like that i believe it was called now no jesus i don't remember label image uh probably you can use the same oh wrecked label so on the mac there is a good one called erect label and uh it's really good i love it and uh i think there are some limits and then need to pay but it's like two pounds or something so i'm paying for 299 a month i think it's totally versus or 1999 one time that's oh maybe i should get this one and not play monthly okay we'll figure that out later uh there are many good tools for image labeling but uh what i'm going to say uh recently i really start to like the rover flow and this is basically like the online thing um where you can like label your image and also it will generate you all your necessary like all these folders folder folder structures and all these annotations and everything so i really like using this because it's just really really simple and as you can see here i already created the keys so if you go here you can see i already uh labeled them uh because it will be extremely boring you to watch me how i do that and uh so but i'm going to show you you can create a new project uh let's call it keys test uh and there is only one annotation group with these keys we are going to use object detection create project in here you can upload your files so for example keys upload this will upload all your pictures i'm going to give it a minute and we will be able to finish uploading then you can this is why i really love this rubber floating because uh it can do all this boring stuff for you uh you can all or of course you can split the images 70 30 or whatever i mean you can you can toggle this and uh but usually the rule of thumb is like something like that i'm just going to leave it to default then it will upload the files and then it will split them and everything you can add some preprocessing and stuff but obviously oh thanks to that they already annotate them you just figure out i already did that so anyway so usually okay that's not exactly what i wanted let's say i remove this until it's enter and how you can add you basically just do this annotation thing i mean it's pretty obvious and then just enter and uh with the arrows you can you can switch between the pictures so you just label the images like this but obviously that's something what i already did so i'm not going to bore you with that i'm going to delete this project okay test and we are going to use the one i already did and 100 images one class only uh just to keep it simple this plate i leave it as default at the moment authoring this good thing if you have images rotated 90 or 180 degrees and resize 416 um i'm going to leave it it's dead now but later on maybe we can do like 640 or should i do it straight away now okay i'm just going to leave it like that and uh let's pre um at pre-processing step we can add some different stuff we can play with it later uh augmentation steps you can do a lot of different stuff if you don't have enough images you can always do something like some blurriness or grayscale or rotation so it can like triple or even more your data sets we can try it later on and here you just click on the generate and we can call that yolo v5 save name and in here it will generate images okay okay and in a minute we should be able to maybe meanwhile i can turn on my camera video capture device okay so i'm back uh no that's not exactly what i wanted but here we go so um you can also start training but this is kind of like premium feature or something and i'm not sure you really can do that at the moment but that's not what you want anyway so export and in here you can see it can export of many different formats if you are using like uh jetsan nano or something you can use the every course uh or like raspberry pi or something and if you have created a mail that's for uh mac xcode we are going to use yellow wifi pytorch and download the c2 computer continue and this is basically what i love because it will create everything for you when i open this up and extract it to let's say let's say to desktop why not but maybe inside some folder called dataset extract okay and when i take a look at the desktop data set and here you have the test images and labels so you don't need to create all these structures by yourself because it can generate for you it spread it 70 30 or however you set it up and also it creates this data what is a uh basically yolo if i requires the images of course and uh where the images are and where the um and and basically how many uh you know what i mean how many how many uh objects we have there right so that's it um what's in here okay this is nothing really so this is what we need right now i'm going to jump back in here and uh well now i need to remember how i was doing that so i'm going to minimize this because this is the important stuff now and maybe i can keep it small that will be better let's just make it a little bit bigger all right now we need to train and uh okay i don't know what to do now so we can take a look and in here there is a training and uh basically we are going to use this train.buy so let's try that so python uh not sure python3 probably train then data and data we have on desktop so desktop data set and it was called data set data of course um data that y aml is the format what it requires and but it was created already so that's fine now we need to say cfg and that's uh basically this yellow weights and if you take a look in here uh i have seen it somewhere um maybe if i search for it um these emails uh maybe i should doesn't matter um yeah exactly so this 5s5m5l and 5 x is a there was one picture of it uh uh let's see uh um and somewhere here there was a good picture what what does it mean but uh if i'm not able to find it quickly i oh here we go so basically this is it these are the parameters like this the small network medium large and x-large uh once i have the really powerful oh it is still running why is this floor i think i don't want it okay why is it still running not sure so i can use the yolo oh it doesn't run again which is up okay so let's copy and paste this again then we can use for example this one so this is the largest network so yolo v5 x right now we need to do weights um but in here i mean it will download them itself i believe so let's just try that and batch size gsi 64. let's try what okay ah of course so just one sec i'm going to look at [Music] um yeah so if i take a look in here in my old ones i see obviously i need to delete them download them so i believe if i go here i can download them so i click in here i don't remember where it was oh yeah okay so i was using yellow e5 x so it's this one so this will download meanwhile i can jump to the directory what is a yellow v5 and put it in here so now i don't have only the small one small network but also the large one because i want to use this one but obviously this one will work as well it is just less complex than this one um and we want to have a good result so i'm going to use that okay uh i'm not going to need this so now if i run this uh okay we have some ah of course i forgot to put this dot pt i believe so let's check again here where was the training oh no like this um well i think it should be pt but let's try yeah exactly i believe wait i'm just confused now but don't worry we will figure it out no just give me a second train okay okay okay so let's do it this way so question three um train now the image should be i have it 416. uh let's leave it like that then the batch uh let's make it 16 then we need to have the ebooks let's make it something like just a 100 for now and the data are at home desktop uh datasets and in here are the data yeah that's what we need and the weights are going to be uh yellow 3 5 x dot pt okay so hopefully let's see what we'll do okay he's doing something data said not found okay but we can deal with that that is enough to found um i think [Music] just give me a second so if i understand this correctly let's double check python 3 train image batch ebooks data oh maybe okay okay i believe maybe i need to copy this or move this data set um into the same folder so where we are here data and in here i can create new folder or actually no i just can do this that will be easier so data set is now inserted data and hopefully this will work now so it's not in the desktop anymore it is a this in here in data like that cle5 slash data slash data set state yeah like that let's try to run this um and i still think we have some issues uh oh i may know okay okay okay just one second um this is something to do with this one because this thing here this uh i can open it here this one dp ai whatever that is it's like you can track your progress and that's connected through the api so i'm going to sign in and yeah so these are the failed ones and so it is connected because uh is it somewhere here trying custom data yeah let's go here okay okay okay we have all of these oh and this is what i was looking for and yeah this is the one and if you don't have this you can install it and this will allow you to run all these progress progress so you can see all the progress but yeah so this is yeah so this is the data set second one second i just need to find it i believe it goes in okay um i think i have everything correct in here let me double check with this uh this ah maybe okay maybe this way maybe i need to run it like this okay that's another solution uh well this is quite shine isn't working but that's how it is always okay so let's double check everything okay let's first try to do this with the s1 to be sure that's not an issue um well this feels extremely stupid thing just one second oh sorry i i just i have seen your question just now hello pablo i have a question for you what are the formulas for determining the ebook size and the base size well um i'm not 100 sure of course but the epochs i mean more epochs mean it's it's trained more really right so uh for example if i run like three ebooks the network will be we can try the differences you will see that there is not it will be no guys great but more epochs mean uh basically it will train the metric more and the batch is only how many pictures it will do at once or uh and that depends on your graphics card so i think 16 was working for me but when i was trying to do 32 it was saying it's too much uh we can we can try that you know in a minute i just want to make this work dataset not found but that's this just oh jesus i know what's the issue okay okay i think i know so let me just um take a look at the old ones um uh uh okay okay okay so i'm going to show you what i'm doing so this is the old folder that's why i don't delete it because i was yeah basically these pads are probably not correct so i'm just going to copy this and maybe i can do it in here so yellow e5 data data set data and like this and yeah this is basically what i want but i need to do this so copy and in here i'm going to do this so basically he just wasn't able to find where the images are i hope that's the issue and like that i'm not sure the top one need to be there once i defined these two probably not but i'm just going to leave it in there i can deal with this save it and let's try to open it now run and yes i think that's it yeah finally so you see you know the ebooks oh yes so basically i just need to update these parts maybe they don't need to be full one they can be something like uh this but i'm just going to leave it like this okay so finally now it's training and you see it's it's really fast but now i selected this small yellow e5 s only what is fine so we can we can see the differences of course that's that's great it will take a moment and once that is done meanwhile week or maybe nothing meanwhile because it will be done in minutes meanwhile i can check if there are any more comments like this no there are no okay okay i'm going to minimize this and here we are oh meanwhile we can take a look in here and here you can see what is happening so i didn't play much with this one but let me just go here you know five no no no oh this is 15 seconds ago so this is it so in here you can see these are the labels but in here somewhere this is the train i mean i don't understand much what's going on here but obviously you can see it's decreasing what's good but i need to discover more later on like what everything means but what i'm trying to find there are these train batch oh this is my old data set uh when i was doing the traffic signs but why is it showing me that i don't want that i want these ones the keys but not here nothing well i was using this like twice in my life so that's why i don't really know where to look for the info but so sorry about that but basically in here i mean you can wander around and look at the different informations but it doesn't matter we are trained now so that was 99 epochs so 100 completed in in the month it's like two minutes or so and now what's happened it save it to runs train and export for now a lot of youtube videos just end up here like that's it but obviously it is not it because you want to see the results so we are going to i'm going to open the new terminal or maybe not wait a second i just want to look for the comment i was using so yeah okay so we are going to use the python 3 detect and now we are going to use our weights which are in uh these in runs or maybe i can show you how does it look like so in home [Music] runs it show in train and x4 so train x4 so this is basically what it spit out there are some learning curves and stuff um some some interesting stuff but i should understand more but i don't really so sorry about that but you can see there are some results well they're great obviously but we want to see them through the webcam because that's what we are um because that's what we want and but in here in weights there is the best and last i'm going to use we can try both but let's say we are going to use the last one so runs train x 4 um and no weights and last okay minus minus image how i was training it 640 i don't remember uh well maybe i can take a look right uh i was training it as well about these resolutions yeah 416 so let's do that so for 116 con this is how confident it should be so we can play around with it but let's say 0.25 and source is webcam so just zero now it will show me the error because i need to turn off my webcam so i'm just going to kill it um yeah i need to turn off my webcam first from the obs okay i'm gone now i can run the same command again and this is my graphic card and here i am nothing's happened until i grab the keys which i don't have next to me so give me a second okay so here are the keys and as you can see it's fine but it's not perfect of course i mean it's fine but we had only 100 ebooks and we have only the small network and the yellow v5s so now i'm going to do the same but i'm going to i mean it's still working kind of fine as you can see but now i'm going to pimp up all the parameters so in here i'm just going to um okay so there's so much of me uh let's say something like this and yep that's perfect i i'm going to do now this more no no i want to use this but i want to say jesus what is this generate new version yeah and in here i'm going to make 640 by 640 because if we take a look in here the yellow e5 x somewhere it was saying not sure where but somewhere it was saying it is using 600 maybe all of the music's 40. okay let's try that so 640. uh i never know if i should just stretch it or do something like this uh we can okay i'm just going to stretch it whatever apply out orient and now if you want we can do some pre-processing steps like no not previous thing but in here the augmentation and maybe we can try something like rotation to both sides 15 degrees or we can try some blurriness and it's too much let's say just like a little bit of blur and then we can try something like saturation okay then we can try something like noise maybe this is too much noise let's do something like that uh we can try something like this that's okay we can try something like brightness that's fine and we can try crop i mean yeah why not but i'm not sure if okay i'm not going to use that at the moment because i'm not sure if it will be [Music] but if it's out of the okay doesn't matter that's enough let's go to continue and generate two i mean pro version can generate more images but we are going to use 240 what is basically almost three times more that's this for free and i'm going to say 640 with oh augmentations uh this terrible name but okay we can live with that um generating images this will take a couple of minutes or maybe not even minutes because definitely longer than before because of all these augmentations so let's wait okay it's moving fine [Music] meanwhile i can try what you can clear this out and to today now we are going to use this larger network with larger resolution with augmentations and i'm going to do let's try like thousand ebooks but i'm not sure how long that will take if it will be too long i will just make it something less but we should be okay uh meanwhile yeah so this is what they were saying you just basically install this vmdb and then it will ask you well i have it already so i cannot show you but then it will just ask you for the api key what you can get out of this website to a new register so that's when you want to see how your training is performing but uh what else we have here yeah so basically these are all the steps for example to create a data set like how does how it should look like and this is the all the parts the images number of classes and then there are the classes but this is what the roboflow will do automatically and the labels you can you can do with overflow as well and everything so okay so here we are these are the labeled images now it's 210 so that's together 240 images exports in yellow e5 and that's that's it and okay download here vr and extract this to home uh yellow e5 data and let's create a new folder 640 with um like an implementation i mean i know this terrible name but we have it already so that's fine now if we take a look inside yellow e5 uh data 614 blah blah blah we need to fix these parts again so i open the terminal in here pwd it will show me where i am so i just copy this close close and i mean it should work even like this but let's say and obviously i have only one object so that's quite easy to train but okay uh okay we don't need this anymore but we can for example take a look at the images [Music] so data train images so you see for example so these are all the augmented images you see so it's it create more images not only once i had but for example rotate its brightness and everything so it's pretty useful to have if you don't have a large enough data set so you can artificially enlarge it like this okay uh we are fine with that also we can toggle these like uh the split can be larger or smaller but enough of that we are going to train so i'm just going to use the same command as before i'm just going to say this is not for this is uh we can try to increase it but yeah let's try first to leave the 16 and we need to s this is not right because now it is in data uh 640 and like that and this one is going to be x yeah and one more time oh and i can i want to increase the number of epochs let's say a thousand hopefully it won't take too much time and yeah we can train and everything seems to be okay okay it's running and here we are so oh jesus definitely i'm going to decrease the large number because this will take quite a long time with all these augmentations uh sorry i'm going to kill you but we are going to do just even 100 now seems like a lot but let's do 100 okay let's try to increase the size to 32 the patch size i'm wondering if this will work because i mean my gpu has like 24 gig oh you see this is out of the menu memory because the batch size i i create it um the bit size is too much so i'm going to decrease it to 16 again uh i'm not really sure what's the logic behind it like what exactly how can i how exactly i can calculate it but should be 16 is working so i'm just going to leave it at 16. uh yeah let's try it again and uh well but this is going to take a couple of minutes now i believe so meanwhile we can do something else okay i believe it will be still less than five minutes hopefully uh but you see now as we have more images thanks to the augmentation and thanks to the fact we have large images not we have now 640 pixels by 640 and also we are running uh we are training it on the uh larger uh network this one yellow v5 that is the the most complex one um again i didn't really try differences between all of them i just tried the small one in the large one obviously the small one is faster but it's not as complex as the large one i mean it's kind of obvious so okay this will take just like five minutes or so we are almost at 10 so at least i hope so so the gpm memory is so this is taking me 15 gig of ram of the gpu ram so that's probably why when i double the batch from 16 to 32 it was trying to use like 30 gigabytes of ram and my card has only 24 and well there's not only this like a lot for graphics card but and you can see my camera is kind of my stream is kind of freezing now hopefully you can still see it properly that's because i'm using the gpu at 100 oh can you please raise your voice and try to make it more clear as you talk uh okay yeah i can try that uh i was looking on the obs and showing me the level of the volume is okay but maybe i'm just talking too fast and with my eastern europe europe access access accent is probably not as great as it should be so i will try to talk more slowly and clearly one second i have just one email i need to reply but we can take a look at this okay so it's 20 epochs so this will take just a few minutes it's not that terrible and one second only [Music] okay i'm almost done great oh i don't want to do that yeah here we are okay about the meetups itself uh well i was trying to contact vodafone if we can now as the restrictions are uh not that limiting i was trying to ask them if we can go back but i don't have any answer so potentially i was thinking maybe we can figure out some different place where we can meet uh i'm not sure what exactly that would be but maybe i believe in newberry there will be some co-work space or some kind of basically we just need a meeting room really so uh i will try to keep looking uh to different place or to try to deal with vodafone if we can go back and hopefully by next time we can do this in person what i believe can be million times better than online so hopefully soon we will see each other what can be great but sorry i want this and we have 35 ebooks okay it's kind of i mean it's not slow it's like terribly fast for the amount of images sizes and augmentations i have because if i want to train this on like some normal gpu like i don't know gt x 1080 or something that will take like much much much longer but the the point why i'm doing this on the graphics card now is because of course you can do the same on the google call up if you don't have a gpu and and it's fine and it will work but when you have like a large data set with a lot of uh with a lot of classes and stuff then on the google collab you have some limited time and there are some workarounds but i mean if it's your gpu you can run you can run it for a couple of days if you want and no one can stop you really you know there are no limits or anything so that's that's great and uh on on the google co-op they will kill your session within like like 12 hours or something i believe so that's why i really wanted to try this on on gpu because there is a lot of time i just need to kind of like track one object and for example with like another options uh for object detection it will take you like half of the day to to make it work but with this yellow v5 it's really fast uh compared to the all and especially when you combine it with this rubber flow i think within 10 minutes or something you can be training so that's really good and uh i mean obviously if there are more classes or anything it will take a longer time but it just it just really works i mean we just install it and within five minutes we was able to start training was great obviously you need to have like nvidia drivers you need to have like the coded cuda installed and cudn but that's not so complicated you basically just download it from nvidia side and that's it and i hope my stream is not freezing or anything because of the gpu i hope i mean as you can see my camera is terribly freezing but i hope that's only camera let me check on my phone and that's because obviously we are using gpu 200 now so uh let's see it's not that bad but yeah you can at least see the real life impact of the gpu because i have only one so with one i mean i'm streaming and training with the same one okay 62 to 99 you know what i'm just going to jump out uh to get some water and i will be back in like one minute okay just a sec okay here i am uh 69 to 99 uh okay we are getting there but hopefully thanks to all this time the recognition will work really well and it's three minutes until eight o'clock but i think we should be able to finish in like 10 minutes so it should be okay uh yeah what else i can okay so if i go to yeah i want to show you the difference between yellow v5 and for example euro v3 because uh yeah here's the nice article about it and how does it evolved yeah first vision second and third and the third one was the last official from the guy who invented it because then he tweeted i stopped doing cv research because i saw the impact of my work was having and then he stopped developing it uh what was in february 2020 but then some another guys the community really created uh yellow before and what was the fork of the original one of course and uh it was faster and more accurate as you can see on this graph so yellow v3 is down here so this is fps and this is accuracy so yellow e4 was really really faster and more efficient efficient than the v3 but then the yellow e5 came and there was a lot of chaos around because it wasn't it doesn't really have any real benchmark and anything but now it's kind of clear it's it's working really well and uh i mean the performance is just crazy and there is this ppolo they don't really know it's some kind of a chinese version it should be even faster especially on stuff like uh which would be faster like on tensorality it was shits and i know and stuff but i didn't try that one because i mean there is not much documentation but the yellow e5 uh is using the pi torch that's why probably is working straight away with any without any issues and uh well i was hoping for the graphics yellow v5 let me see if i can google it uh like the performance difference between something like this but i want this but with your ob5 no well i mean even here okay we can see the yellow e3 this is not exactly what i wanted uh vs eo love e3 and um not exactly what i want but yeah i mean you get the point he's just he's just faster this might be no this is sweet tree i mean yeah whatever you get a point let's let's see ah now we have hundred epochs perfect so we can just you see now it's just uh create this last and best maybe we can try both of them of course if there will be any significant difference well i doubt it because uh probably will be not able to figure it out just by i um i forget to take a look at these uh jesus how it's called v w or something this uh yeah this is what i want so login oh something happened uh well okay so this is the one and here we can see the performance um so these are the labels so this is how it's labeled and these are the predictions so it's pretty good but i want to see some stuff like this uh okay well maybe if i leave it for longer but we don't have time for that uh because last time for example i was trying i will i left it for a couple of hours for 3000 epochs and i mean that was awesome of course but let's try this one so detect but we are going to do 4.7 and let's do 640 this confidence i'm not sure i'm just going to delete it and let's see without it we'll do it we'll put some defaults what is probably the same as 0.25 or ah of course my camera again i need to kill it bye bye and here we go again and we'll see what it will do maybe i kind of overkilled it with with the okay it's pretty good uh but i don't feel that much difference from the last time for this venus but or maybe i do well if i leave it for even more epochs probably better uh as i'm streaming it's not 100 uh the fps is kind of dropping a little bit but i believe that's because i'm streaming as well but let's say am going to use i just want to kill this and i'm going to use the best instead of last and what that will do yeah i think the confidence is slightly higher in this one and yeah i think this is pretty fair enough oh we can try something like this yeah i think he's nice he's working pretty fair okay i didn't expect this well obviously if i really want to do a key detector i should in my data set i should have also different type of keys but let's say as i have one in these ones that's that's enough for the test and as you can see the fps is crazy fast it's perfect but that's because the graphics card is kind of overkill well no does my head look like keys it's not good but not sure what's happening with my head at the moment maybe if we do the threshold uh let's go back to here train um not exactly here's the this detection detection this no no that's not the one anyway uh i believe let's let's just put it back and uh not sure what was that with my hand but uh maybe if i add more images what is not the case then obviously it will be better but you can see how in few minutes really you can create your own object detection with the yellow e5 and it's working pretty well obviously if i add more images or if i um just train it for a bit longer or something it will work better but i think it's pretty pretty fair to say it's it's really nice and okay so one second resolution is like this okay again i'm going to make myself smaller or maybe i can leave myself like as this so thank you everyone for attending today i i just want to keep this short short is one hour or so so hopefully as i said by next time we may do something hopefully by next time we can see each other i know i keep saying this about you know how it is yeah okay uh by next time maybe we can try something like semantics in the segmentation or something like that and yeah see you next time then bye

Channel: Pavol's Lab

Views: 3,410

Keywords: yolov5, yolo, machinelearning, ml, ai, artificial intelligence, machine learning, yolo v5, object detection, deep learning, finetuning, data science, python

Id: zDcS-lmbmuk

Length: 67min 5sec (4025 seconds)

Published: Fri Aug 06 2021