Everybody.

And welcome to an update to the Python police.

Siri's in this video.

I'm gonna be going over just some of the updates that I've found with the object detection and kind of what?

What I'm thinking.

Moving forward.

So this is one of the out of the box Coco models of detects cars, pedestrians, traffic lights.

Although there were really no traffic lights around here for me to show you.

But it's pretty good.

The problem is, it still just doesn't doesn't detect as far like the distance wise parted plant over there in the corner isn't quite detect out as far as I would like it to, um and then also, like, see that car still not being picked up.

Really?

That card ever got picked up like the very, very end.

Um, and really, all I I want his cars in people to be picked up is good as possible.

So probably gonna wind up having to train my own custom model just because this one's not picking up people as quickly as I would like.

So the plans for the on foot model or tow, you know, play g t a in the way that one place cheats?

Yes.

So he should be able to engage with the police here.

And the way that we can do that is you know, when you find out Hey, where's the police or whoever you want Thio engage with you.

Just find the center of the cube or square Rather, and that's how you engage now.

Same thing with stealing cars the way that we did it before.

If you would find the bounding box for the car and then basically you would move the mouse to force the car to be basically center screen run towards it.

And then when you're close enough press f to enter the car, that's what we had done before.

And I think that's probably what I'm gonna do again as faras like stealing cars is concerned.

Um, but I think would be cool to let, uh, let the Aye, aye, kind of, ah, do its own thing on it on foot rather than purely go for cars all the time.

What did that say?

I'm not sure what that was calling that yellow line.

So, uh, so with the models, all I really did was just Ah let me see if I could get away here.

Yeah.

So this is the Grand Theft Auto object detector.

Basically, we're just grabbing Screen is use rule.

I do also have.

So it's also it can run in both at the same time.

I just set the GPU fraction Keep clicking on this toe to play.

Um, so I can unplug is And that's the other thing is how do we get Charles toe like, walk around in the environment.

So on the one hand, when he's in third person, he actually can walk around pretty well because he's just gonna walk is if he was driving.

So he should pretty much always walk along the roads, which should provide decent enough play.

The problem is, generally if you walk, the police are always gonna appear behind you.

So it's kind of awkward, So probably still gonna have to figure out some way to engage with police.

The other option that we have is the on foot version could be completely retrained, using similar methods that we used to train.

You know, the self driving car on Lee.

I'm thinking that you could probably do transfer learning on foot.

So start with the car model and then from there go to do transfer learning.

And I think we should be able to do that.

Then you have some sort of, like state that you're just tracking.

Like, where is the player is in a car.

Is he on foot?

And then later, you know, Is he in a plane or all that stuff?

So, like, as you can see, pretty, pretty well can just walk along the street.

But yeah, you're always gonna have police behind you, which is kind of weird, but they always do seem toe end up being behind you, which is kind of lame.

So anyway, I gotta do something about that.

Maybe you could just use open CV.

Figure out where the police are by looking for, like, the little dots or something like that.

Or you could train a neural network model to always, like, kind of turn towards police.

So no, not really any different than how we trained the self driving car to kind of follow the purple line.

I don't think it would be too difficult to get a aye aye to, you know, look at all those like red or blue dots and try to always aim towards the red or blue dots, so that should be totally possible and hopefully require a lot less data if we're able to get away with using transfer learning because it's the same exact he presses that we'd be pushing.

Um, And then you would use the object detection script itself to actually when you when it comes time to aim to aim.

So, um, anyway, I think that's kind of my ideas for how we're going to get a decent on foot version, but we'll see how that goes.

Surfboard nights.

Nope.

Anyway, um, yeah, So, uh, some of the ways, like, if you wanted to play with this yourself anyways, uh, you could come into here and all you have to do is if you pull up, let me just bring it up real quick.

Um, it's like the GSR GT tensorflow object detection ap I model zoo.

So that t key is for pause.

So that's what happened there.

Um, really the get help.

There we go.

So if we scroll down, these are all the models that we can use and then here is the milliseconds and again I still don't actually know unfortunate and published the other video before filling this one.

So someone's going to be able to tell us hopefully at some point, what what this is.

I don't know if it's like a score or what.

This is definitely how quick it runs, though.

So the one I'm using I want to say is probably like maybe this one, the or CNN inception.

Let's see which one I'm using.

Yeah.

So the RCN inception V to Coco.

I tried some of these other ones, so this was the best one I could find.

So there's always this kind of, ah, you know, a balance between accuracy and the speed.

So, like, for example, some of these air really slow like this is 600 milliseconds, so way to way too slow.

Um, you know, really, even like 100 milliseconds is is just too slow.

We need something faster than that.

So, um, anyway, obviously the faster the better, especially because we're trying to find that bounding box and then in theory, shoot towards the center of that box.

Well, if that box is actually moving in game, then you're probably always gonna be like lagging that box, so that'll be a problem.

Um, Anyway, you can try all these, and if you want to, you can't just take the name.

But if you right click and copy the link address, and then you come down here just anywhere really paced it out because they all have, like, dates to them.

And then this is the thing that you want to pass here.

If you don't do that, you'll get like a 403 Forbidden.

So anyway, so you can try throwing those in there.

Um, this is basically the exact same script as in the tutorial that I've already done.

So I'm not really lips.

I'm so used to being alone.

Just type python.

Um, it defaults to python program that net g t a.

And if we scroll down Thio object detection, it's basically this script here, So really, the only difference is just bringing in the newer version of object.

That section in a different model.

You could still try out this model, but this is quite old.

So you definitely want to use the newer versions if you can.

Um, I think that's it.

So that so the next thing I'm gonna work on is, um probably a transfer learning a pride transfer.

Learn from, um, from this model.

Since this one seems to be the best one.

But I am curious to try something more like anything that has, like, the slows are the quickest speed.

So I don't actually know.

This one looks pretty quick.

26.

I don't know if anything is quicker than that.

26 milliseconds.

So, um, so, in fact, what I could do is copy link address, pay whips.

What happened here, Paste?

Uh, come over here and we could paste in that mobile net.

And then we come down here, break this, rerun it, and we can just see, like, how quickly does everything run?

The other thing I'm not sure about is, like, right now, this is a 12 80 by 7 20 but I can make this size anything I want.

So, like, for example, toe before I send it toe process, I could make that a, you know, for 80 by 2 70 or whatever sized image.

Um, so did you see this is much quicker than the other one.

Although it doesn't seem to be there, coz Yes, Oh, look how like the frame rate on this is pretty good.

So the one that's top left, this is the one that's actually running.

It's pretty good, but you could see it doesn't have any distance to it.

So, like, the other model was would definitely have detected that cord Private.

That pedestrian or the police guy walking there.

Look how close I got to get to this car were close.

Oh, my gosh.

Come on.

I can't believe it doesn't get that car or the pit it.

Okay, so anyway, they're sort of goes.

So this one, Unfortunately, it doesn't quite work.

And I'm not totally sure why, Like, maybe the resolution is is too high.

I'm not really sure like you.

Maybe if I bring it down.

Um, because I think that matters like how big the input images I am kind of surprised that we're not detecting that car.

Let me.

Let's break this rule quick.

I'm just kind of curious.

Um, some of this stuff I have the answers to as I'm filming and some of the stuff I really don't.

Um So if I come down here, um, I want to say OK, so this is so I'm gonna try for 80 bye to 70.

Oh, and then when I pop it up, I'm gonna do is I'm gonna throw in a re size so CV to die resize so just before it shows the image Because 40 by 2 70 is gonna look really, really small on this screen because this is 12 80 by 7 20 and you can see it's already quite small.

So for 80 by 2 70 is gonna be, like, super small.

So see what you got resize got rid of mine.

Quote there, Um and then you re size 00 and then the function that we want to resize for exist just two times and then function.

Why?

It will also be a two times going to see you that come over here, run it and I'll come back in here.

We'll see if it detects that car.

Hopefully the mods pregnant.

Teleport me any second.

Now just wait a minute.

Not sure I don't think escaping helps the mod.

I think the Mahdi's is its own stand A long time, so we'll see.

Let's see if right at the end, it gets it.

No.

So I don't think that's it either.

As faras Well, we way.

Think you're me Pause this real quick.

Let's see if I go back over.

Yeah, so I think this model just see, that's the unfortunate thing is that it's, um it's fast, but, uh, it's not detecting the things that we want, but we may find that via transfer learning it does on.

We can get that this frame rate, which is, you know, I wouldn't want to play it this frame rate, but I can, you know.

So anyway, those are the things that we're gonna have to definitely be checking out.

But like I said, the best one that I've found so far has been, um, this one here.

So probably all you would do something like this.

Like when I go to training custom model.

My next step is going to be a training custom model transferred from this model here and then maybe later transferred from, like, this one.

Just because this one runs so fast, I'm actually just gonna make a note of this myself.

Runs detects poorly pieced.

Okay.

So anyway, um, those are the nice sex I'm gonna be working on, Uh if you've got questions, comments, concerns, so just runs.

Feel free to leave them below.

Hopefully, at some point, someone will tell me what?

That already closed down.

I didn't hear what this is.

Coco M a P If I could google it, what does that mean?

What is the cocoa?

Here we go.

Could someone just help?

Can we just be told what it is?

So it's like speed the accuracy trade off maybe.

But is a high score good or bad?

Okay.

Anyway, um, I guess I have to read about that later.

Someone can comment below, though, Like, this is the one that we were running out of almost wonder, like speed divided by like, if accuracy was high, then this would be very low.

So I'm not really sure what we're looking for.

To be honest.

Anyways, um, that's it for now.

That's what we're gonna be winding up heading towards.

The next thing is just getting the aye Aye to play well enough on foot.