Comparativa de Capacidades de Visión AI: Jetson Nano Super vs Raspberry Pi 5

when it comes to locally hosted AI my interests tend to lie more around running llms or perhaps stable diffusion

or even video generation and things of that sort however I really can't help but notice that there is a lot of

interest in the actual Vision capabilities of AI capable devices now when I say Vision capable I don't mean

running a multimodal llm which can see a picture and answer questions about it but I'm actually referring to the

ability of little devices like these to actually take a video stream as input and be able to detect and classify

specific objects from within that stream now that essentially leads us into the whole point of today's video which is

going to be a sort of fusion of a few different things one being a kind of tutorial a kind of comparison test and

just a kind of benchmarking test between the vision capabilities of the Jetson or a nano super

and the Raspberry Pi 5 with AI kit to actually go ahead and run some of the software that will allow us to

experiment with these image capabilities is not as simple as just downloading olama on one of these devices and then

running inference on an llm you do need a few extra things and that is really kind of why I have this little orange

layout or spread if you will in front of me so of course from right to left we have the oron Nano super which if you've

been watching my channel you're likely rather familiar with at this point as am I the pi 5 8 GB we have a camera cable

which will actually allow me to plug in a small little webcam to either of these devices we have an opened but unused

Raspberry Pi 5 AI kit now it does seem that they came out with a more powerful one of these however mine is the up to

13 tops version so just to make note of that I have an opened but unused extremely cheap 5 megapixel webcam from

micros Center or Inland which is the Micro Center store brand and finally I have a Raspberry Pi 5 unopened unused

active cooler now I do need to go ahead and actually install the AI kit and cooler onto the pi5 so I'll probably

just timlapse that as I don't really want to spend a lot of time showing you me like screwing things into a circuit

board and then following that we're essentially just going to start with either of these devices and go through

the steps to install the necessary components and start to R run and see how they

[Music] do we're going to begin with the pi five because I put the camera on that so I

figure let's start with that we're now going to go ahead and just follow the steps to get a basic pipeline running

which essentially as it states here will demonstrate object detection human pose estimation and instance segmentation to

begin we're simply just going to um not do that okay to begin we are going to just

clone this repository once that is completed we will navigate to the directory

and once in there we will run this installation script once that has completed we're

just going to keep going down the line here and we are going to configure the environment by just copy pasting The

Source setup env. SH now that that has been complete we can go ahead and do the PIP install d r

requirements.txt which we'll go ahead and installed the required requirements in this

environment and it says the rapid json. deev package is typically installed by default on Raspberry Pi OS I have not

seen any indication here that that package is missing however I am wondering if maybe I'll

just try this anyway because if it's already installed it will just say the requirements already satisfied okay so

that's fine maybe perhaps just a good thing to check I suppose we are going to download the required Resources by

running this script and basically from what I am seeing there it looks like it errored

but it just seems to show that these required resources are already downloaded so it's not going to retrieve

them so I would assume that this is actually okay and following that we are going to run the postprocess compilation

script well that kind of sums up the initial seems like preps to actually go ahead

and start playing with some of these examples so we are now just going to run this example now considering we are in

the same terminal session and have already done this we don't need to do this again and we are simply going to

just go ahead and run the detection example now I've have not done this before I do have the little webcam kind

of plugged in and hanging off the table so I would assume that hopefully what we see from here

is oh okay so my mistake but this is actually this is actually really quite

cool because it is working so it is running it on the demo video which is what we see here but the bounding boxes

along with the actual labels and percentages on each are actually being generated from the Raspberry Pi's AI

accelerator so this is actually really quite cool and we can see that it is just kind of doing its thing I I don't

know why but like for some reason I want to like touch the AI accelerator to see if it's hot I'm not going to do that but

this is actually kind of cool so now obviously we want to do this

with our own video now if you are using the attached Raspberry Pi camera as I am essentially we're just going to click

the running with different input sources Tab and that will bring us down here and there is actually a rather simple

command that just adds on a flag of input RPI which should essentially just be the default Raspberry Pi camera

plugged into one of the camera ports so we will so what I'm seeing here does not

look good are just going to try this again once more with the Raspberry Pi camera input and we will find out if the

error was on me or if it isn't yeah okay so all right so here's a top tip for you folks if you want to get this working

and not look like a the gold contacts on the camera cable need to be facing towards the ethernet port or all

of so the gold contacts here need to be facing towards the ports on the pie and not away from the ports I'll show a

picture of what I mean so oh wow this is actually working right now even though it's upside down all right I probably

will end up using this cuz this thing awesome and I bought it cuz I was enamored by its miniature yet Mighty

look but basically we're just G to go ahead and oh wow this camera is not that bad person 90% so hey we're running

right now ai object detection on the little pie chair uh not quite all right what are you looking at up there CU

that's a laptop oh I got it look at that look at the the sorry I got excited the iBook G3 is being identified and on top

of it okay so I'm sorry this is probably terrible right now because I'm not really umbrella I could see that I would

I accept that this is really quite interesting and this is honestly the first time I've actually played with any

of this truth to be told I won't say that I I didn't find it interesting but I just couldn't really think of many use

cases that would pertain to my own personal workflow with this if you will but this is actually really cool so it

thinks this red Keys a boook all right here's what I'm going to do I'm going to go ahead and get the camera situations

sorted so that I actually have a stable frame of reference for this and then we'll just try having it detect some

things but this is really cool so if you're running a script like this and you want to get out of it in the

terminal like this you just press contrl and c and that will gracefully end the process very cool after learning that I

have absolutely no idea what the correct use case for a telephoto lens is we are now just going to revert to using the

simple little camera that I had origin originally set this up with now to get back into the examples if you don't want

to have to go back to the GitHub in the terminal you can actually use the arrow keys to cycle through previously entered

commands so essentially all we have to do is just go back into the directory we need to Source the virtual environment

and upon completion of that we basically just run the python script again with the flag for the Raspberry Pi's camera

now you can see I do have a few things in here we have a mannequin over there and it is actually detecting that as a

person which is pretty cool so now um don't mind me CU I have actually just taped the camera to the monitor so if I

uh kind of like use the monitor as as tilt so it's showing me with pretty high detection again I'm not very versed in

this so I figure we'll just kind of put some things into the screen and kind of see what's going on so um like a an

automatone okay it thinks this is a remote but all right um person what about let's

try an Oculus headset apologies if this is very difficult to watch right now as well as here because I have to kind of

keep the mic with me but let's uh remote okay so and again you have to remember that this is only using the

training of the models that okay sorry what am I trying to say the model is only trained to detect specific item so

it is likely if you were to train it to actually understand what a quest headset is or something like that then it would

absolutely be able to identify this as well but it seems this is more for like testing a person or something like that

you can see that it does show the laptop behind me well it did before we saw that originally when I had first actually

tested this so this is interesting and I suppose I'll try it with one more like interesting thing so I'll be right back

I need to go and acquire a dog all right so I brought a dog this is Betty she doesn't like it up

here but it is getting dog with 90% accuracy so really kind of cool and

yeah okay good job as a quick side note I've been making a gaming computer well a micro gaming computer case for the

Orin Nano so I just I won't make a video on this but I just wanted to show so if people are interested in a case I'll

have this finished at some point um and it's totally not inspired by anything at all we now have the Jetson or a nano

super setup the camera is plugged in and just like we did with the Raspberry Pi we're essentially just going to follow

along with the tutorial on how to get the object detection library and model running now the Jetson is a bit more

complicated for the initial setup and actually getting all of this to be able to work as simply as you will likely see

it work right now I did put out a rather in-depth tutorial that if you are interested in doing this on the Jetson

would likely be a good place to start before you just try to jump in and kind of follow along with what we're doing

here now with that kind of disclaimer out of the way I am basically just going to go ahead and run this command which

will essentially download the container that will allow all of this to run on the Jetson pretty simply fingers crossed

hopefully and this will take a little bit of time so we'll fast forward and once this is done we will hopefully be

able to easily play with some object detection once the container has finished downloading it will essentially

drop you into the shell for that specific container now I have just gone and exited out of that because there was

a few other steps that I wanted to go over before actually trying this mainly just kind of running through this one

myself first as sometimes there can be a few more intricacies with getting some of these examples working as opposed to

the one we looked at on the P now being that I have done that there are a couple of considerations I want to quickly

mention and that is here it just says for the live live camera example ensure you have a camera device connected

however if we go into the Jetson containers GitHub the instructions are slightly different and they specify you

need a USB webcam device connected for this example now I was trying this with the Raspberry p cameras and was not

having any luck so I eventually went to the GitHub because sometimes the instructions can have some different

verbiage and things like that which may be able to help you more on the GitHub and I tried to USB web Cam and it worked

for shot so basically all we're going to do now is just go ahead and run this live

example and basically we'll get a bunch of terminal output and things like that some things

that may look somewhat concerning however ultimately we will be met with a URL which we can go to and we will see

the example running locally on our Jetson now we see our URL and you can simply hit control and then click on it

which will open it in your web browser however that would open it in chromium for me which I don't want to do I want

to stick within Firefox so we just going to do it here and sorry it's hard to like keep the

this all oriented but we can see so we see the actual image there but we don't actually see any bounding boxes or

object classifications which can be a little confusing on First Site So eventually um essentially for this we

actually want to specify what we have this this looking for so I'm just going to start with person and we put brackets

around that and that actually will go ahead and then draw a bounding box around it with the classification when

it does see a person so I actually saw for a second there's a Napoleon Dynamite I understand now why like weather people

have a difficult job because they can't see the Napoleon Dynamite right there was quickly showing person for a second

and you can see right now when I point to it that seems to somewhat help it so we can basically just I think we can

stack these but I don't know so I'm going to try to oh wow all the laptops showed up when

I hide so this is very cool as you can see definitely this is a different sort of

implementation than what we saw with this working on the Raspberry Pi but it is this kind of the same thing at the

same time so I'm going to basically just uh try like I'm want to try

to let's see I don't know what this is trained on so depending on how these models are trained

really allow me to rephrase that with an llm or something it has knowledge based upon what it was trained on so say an

llm was not trained to know what a VR headset was then it would have no understanding of what that was so this

is essentially the same thing so I want want to type VR headset but I understand it very likely will not know what a VR

headset is because I don't know that this would have been in it's okay I'm I'm not ready to give it

credit yet because I want to do one more thing I want to see if I can trick it to see if it was just kind of guessing

there and it was a lucky guess um okay so if it says these are sunglasses then I'm not

impressed okay even if I I mean I can't see when I okay so that tricked it but if I put

them without my face it doesn't so this is quite interesting really and as this is one of my first real experiences with

these sorts of Technologies I am impressed with this and now that I actually am using it

Hands-On I can kind of think of some more use cases and things like that for something of this sort I do want to show

you one final thing and that is essentially that you can browse to this from other computers on your network so

I'm going to hop over to this one real quick and basically we'll be able to see the same interface on a different

computer which could be of use for a lot of folks have talked about home automation and things like that so

basically you could have an edge device like this set up somewhere in your local network around your house or something

like that and then you could just kind of hop onto your computer and monitor it or have it send you alerts if it detects

a specific object that you want alerts for or something of that sort so it is rather cool I guess to say the

least oops and it's so it's hard to see and hear but we can see on that computer

which I'll put the screen recording up from this computer for a second but you can see that it is the same sort of

thing so just kind of cool and the local network capabilities of this are very interesting and you can actually see

here in the terminal output on the actual jets in that it's saying hey um 125 is accessing this on Port 7860 on

this machine 125 being the IP address of that computer so overall this really quite cool and again this is just a sort

of brief light touch on some of these things and not really an in-depth tutorial as I want to get my feet wet

with some of the image stuff as well so we'll try I suppose one more thing um let's see if it sees that that

base over there I'll just call it a guitar to make things simple but yep very well that's very

interesting let me try book because I do have some books stacked up on my AC there and okay not quite it seems

like the chair is kind of getting in the way yeah well I could probably play with this for longer than you would want to

watch me on camera just do this so to wrap things up this was just sort of meant to be a sort of comparison but not

necessarily A head-to-head performance comparison but just kind of a comparison of the overall ecosystem and steps you

would need to take to actually go about using some object detection on either of these little AI powered or AI enabled

Edge devices there are some considerations when you're thinking of either of these namely price now this is

a pi 5 8 gig with the AI hat and this is the lower powered version of this AI accelerator overall when you factor in

micro SD card power adapter Etc this is going to run somewhere probably around $180 for everything you need to go to

Plug and Play not including the camera and of course the Jetson Orin Nano super is

$250 and then of course you also need to just make sure that you have the micro SD card for this as well and if you want

to put an SSD in it you can um but don't I haven't gotten to actually boot directly off the SSD on this yet I know

a lot of folks were interested in that that is something maybe I'll look into in the future and do a tutorial on but

for now I'm honestly okay with how the micro SD has been performing and basically this was just

kind of a cool test and this was one of my first real Hands-On in-depth experiences and actually using the image

and object detection classification capabilities of AI Edge devices and I have to say this was actually kind of

fun and I am going to ask those of you watching to do me a favor and I'm not going to say subscribe I don't do that

if you can think of a good use case for this Raspberry Pi telephoto lens preferably in conjunction with the

example we saw running on the pi with the AI kit I really would like to actually just have this running

somewhere like 24/7 along with this and seeing what it can do because this really was interesting to me and I

unfortunately could not get it working enough well enough to show um the image that I got out of the telephoto lens was

just super blurry and I attribute that to my lack of camera etiquette or knowledge so that's essentially going to

wrap this video video up I hope it was useful and helpful and happy new year thanks for watching