An actual Q&A for my video about actually tracking Stealth Fighters with cheap webcams.

Thanks for the life-changing support in

my last few videos, and sorry for the

double remaster. Sadly, YouTube doesn't

let you edit videos, and it was meant to

be my summer of math exposition

submission, which I was hyped for, and I

didn't know how to fit this into it. As

a probably necessary recap and summary

for those who haven't seen the series

before, which you can feel free to skip

using this timestamp, I made a video

where I introduced a technique which

allows you to detect and track faint

objects such as drones, asteroids, and

stealth fighters, far better than any

other technique in existence, allowing

you to detect and precisely track them

from well over a 100 km away, which it

achieves by first extracting the motion

values from a series of images and then

taking this motion and back projecting

the motion values. of each pixel in the

direction of its origin into a grid of

voxels and adding the motion of each

pixel to each voxil the ray collides

with and then repeats this from multiple

different cameras which allows us to

average out and filter out noise and

objects such as insects and birds from

the foreground leaving just the objects

that are moving within our grid. This

technique can also be used

interchangeably at night from well over

a 100 km away using thermal vision and

with radar to majorly improve your

ability to see through clouds, not to

mention solve a whole bunch of other

problems. With that out of the way,

probably the most important question is,

doesn't this already exist, though?

Well, that's because there are plenty of

techniques that on the surface sound

exactly like they're doing the same

thing as this, and they can perform

almost all of the same functions, but

only on high signal to noise ratio

targets, at which point they're just

doing basic triangulation. A major

mistake in my videos was that I wasn't

very clear about what the major

fundamental limitations of those

previous techniques were and what this

is meant to be the best at, which is

specifically detecting low signal to

noise ratio targets in cluttered spaces.

as a brief summary of the major weakness

of what was previously the best-in-class

technique for tracking faint objects

like this, which was also easily the

number one most commented technique by

you guys under my last video, which

sadly doesn't really have a single

unified name. So, I'm going to have to

sum up most of the names that get lumped

up into this as being multi-ensor

tracking before detection triangulation

using calman filtering on radon hull

transforms. The problem with this

approach is that it always inevitably

runs straight into a brick wall of

performance issues on ultra low signal

to noise ratio targets as it relies on

being able to call off false positions.

And unfortunately, when your object's

motion signature is faint, you won't

have enough signal compared to noise to

decide what pixels in your image should

be called off and ignored from further

processing. This very rapidly becomes a

major problem as each extra point on

your image that you're unable to

confidently cull adds a ridiculously

exponentially accumulating amount of

extra associations that you have to

check which will cap out even the best

computers before they can process more

than a small fraction of the pixels on

two cameras let alone five or six which

makes this basically useless for ultra

low SNR. But the real killer is that

each extra camera you add increases the

power to which the total candidates that

you have to check is raised to. So it's

more or less computationally impossible

to see any of the gains that you can get

from just adding more cameras until it

works. Which is why you pretty much

never see systems using more than one

camera. And if you want an incredible

video about the difficulty of finding

planes, even using telescopes, then

check out this incredible video by DST

Studios on the topic, which ironically

was what inspired me to realize that the

asteroid tracker that I was working on

could do this. This is why historically

it was pretty much never used in

practice for low SNR tasks as a small

amount of juice really isn't worth the

squeeze because it was pretty much

always going to be cheaper to just buy a

higher quality camera instead of dealing

with the many many headaches and

impossibilities of the older technique.

But because of this new method, you're

now able to add all of your data into a

voxal grid first to combine, correlate,

and boost the signal to noise ratio of

candidate pixels inside voxels, which

otherwise would have been called off,

which makes it much, much easier to

reduce the amount of candidate pixels

with each added camera, all while only

having a linear increase in

computational cost for each additional

camera. Especially because each camera

you add exponentially increases your

ability to differentiate your target

from foreground and background objects.

So if you don't have enough signal from

one or two cameras, you can just keep

throwing more cameras at the problem

until it works. Which is why this has

always been a theoretical dream

technique. Obviously there are a lot of

things wrong with that summary and it is

a bit more nuance than that but we would

be here all day. So broadly speaking

those really were the limitations with

previous techniques. you were either

sacrificing precious speed or

sacrificing crucial information. Which

is why pretty much all of the work you

will find on improving multi-ensor

tracking before detection triangulation

using common filtering on radon hold

transforms was just focusing on the

smallest gains you could get from better

culling. And that's not me trying to

knock on those other techniques as they

not only can but absolutely should be

used in combination with this to unlock

ridiculous amounts more data which is

especially true for enhancing 3D

scanning techniques like cryo electron

tomography where you can make very

reasonable assumptions of certain

substructures appearing in your sample

to help you determine its orientation.

And I've never been able to find an

example of this, including Andrew's

lattis system, that was able to show

results anywhere near as good as I was

able to get. I have tried my absolute

hardest to find any example of this

before me to credit them. But even the

extremely thorough Center for Strategic

and International Studies report on

using distributed infrared sensors to

track stealth fighters only ever

considers the case of triangulating

stealth fighters once you've already

found it. Moving on to defraction

limits. While yes, if you're using

thermal cameras at nighttime, defraction

does start to become a problem for ultra

cheap thermal cameras. Cheap is of

course relative, and it's still well

within the limits of already existing

thermal camera systems while still being

both cheaper and much better than any

other technique. The use of webcams is

just to show off its power. You would

always want to use better quality

cameras to eliminate headaches earlier.

You also gain the added benefit of being

able to use the information from

multiple cameras to create a very

accurate firmy estimate of the position

of the object which allows you to

overcome defraction and other blurring

effects such as photon shot noise etc

etc. In terms of improving radar to see

through clouds better. This also helps

and uses a completely separate technique

to multi-static radar which has been

done for decades and is what you're

thinking of if you've heard of

networking radar dishes before. I would

also expect that radar would see a

drastically bigger boost than what

infrared gets from this as radar

contains a lot more data to correlate

such as distance and velocity while

infrared only gives you brightness.

Also, pretty much all of the problems

that you can think of for this usually

have surprisingly interesting and

reasonably easy to implement solutions

if you throw a little bit of brain power

at figuring out the solution. And even

for more extreme applications, they will

still be well within the scale of what

you can do if you have a few million

dollars to throw at the program to

develop this, which is still very cheap

compared to what this can do after you

solve those problems. And all the

hobbyists who replicated their own

version of this said that they were able

to overcome most of the problems that

people including me usually think of

with this. While I don't think it is

productive to measure the success of a

technique that can be used for plenty of

other things against a series of

programs that cost hundreds of billions

of dollars in total, even if this does

make stealth fighters entirely

ineffective at being stealthy, I still

wouldn't expect the contracts for them

to get cancelled for at least another

year or two, as I would still want time

to do feasibility studies for

workarounds and alternative use cases

since technically if you are able to

knock out enemy radar, you could still

be invisible inside cloud cover. Not to

mention that it's free money to those

who actually matter. Anyways, that being

said, given that in order to be within

the range to drop GBU57s

during the Natan strikes, the B2s would

have had to travel within 20 km of the

facility, which also happened on a night

that we know had clear skies, which

means that you can use the background

stars to plate solve this orientation

stabilization. I do wonder if an

unintended consequence of me releasing

that video, at least in part, was the

sudden unexpected rush to bomb the

facility before a minimum viable

tracking system could be developed. But

I suppose we will have to wait for Ken

Clippenstein for that. Moving on to

things that are actually useful. While

without information about the phase of

the wave that you are measuring, this

doesn't give you the 2D ultra resolution

that you get from intererometry.

We're also only able to perform

interferometry with synthetic apertures

bigger than a few dozen meters on large

wavelengths such as radio waves. So, I'd

still say this is more than pretty good,

especially for asteroid hunting because

it allows your sensors to be as far away

from each other as you want with

whatever wavelength you want. and it

still gives you a ridiculous amount of

extra free information specifically

inside 3D volumes that you don't get

from all of the other techniques which

don't use difference images as well as

the normal images and instead just use

the normal images which I explained

better in my third video. But to expand

on what I said in that video, this uses

information that isn't accessed by other

techniques for 3D scanning such as radon

transforms and in particular gausian

splats and nerfs, etc., etc., which are

much better at a lot of other things.

This is just better at other things that

they weren't designed to be good at or

needed to be good at to begin with, as

despite their incredibly high visual

fidelity, they still struggle at finding

faint small objects like this because

why would you need to? You would never

notice them anyways. And it's also

really cool seeing how they work

different to my technique behind the

scenes. People have also told me to

patent this, but that ship has sailed.

And let me know if you want me to upload

a full translated version of what is

probably my favorite replication of

this, which is an entire de vlog series

made by a fan from China using cheap

webcams. And if you've built something

with this yourself, then feel free to

send it to me by email or on Twitter. As

if you guys want to see it, then I want

to make a compilation of submissions.

Speaking of which, something that I

really didn't expect about having a

video blow up is that pretty much all of

the people who send you something send

it within the first 30,000 or so views.

After that, it falls off like crazy. As

in within the first 10,000, I will have

people within my city messaging me. So,

don't be afraid to send something

regardless of the view count. The

closest paper I could find to this was

this one, which comes painfully close to

doing the same thing, but unfortunately

doesn't quite put all of the pieces

together to do things like accumulation

of ultra low signal objects and is

instead about finding the whole of an

object that is already high SNR, such as

a person. The next closestish study that

I could find was the 2017 pyramid scans,

which use a similar back projection

technique, but don't use a pseudo motion

between sensors on top to improve their

resolution. despite already having close

to ideal spacing between the scanners.

Anyways, I want to thank you guys again

for your incredibly life-changing

support on the last few videos. I'm

planning on going back to my normal fun

videos in my spare time now that my life

has settled a bit. So, we will be back

to your regularly scheduled

deprogramming unless Google wants to

partner on a video to show off how their

Tekken platform enables progress or

something. seeing as I posted this on

their platform. Although, the thing that

I want to make is so cool and

interesting that I might just make it

anyways for the fun of it. To recap,

this improves on previous techniques by

allowing you to find ultra low signal

targets, which would previously cost an

impossible amount of computing time to

find and wouldn't let you keep on adding

more cameras. It works at night from

over 100 km away using thermals and

through clouds with radar, and it is

separate to other network radar

techniques. It isn't phased by

defraction or photon shot noise while

still being cheaper and more than

effective than alternatives. It can also

be used to extract more information from

3D scanning data sets provided that

there is enough parallax between the

cameras for the objects to appear in

different spots. So, it won't help you

with 2D scanning like taking images of

distant galaxies which is what you need

interferometry for. Even though I tried

to go over some of it again, there's

also a bunch of extra stuff in the third

video that I don't cover here. So, some

of your questions may have already been

covered there with most of the extra

stuff happening at these timestamps. But

yeah, it takes ages to edit videos and

this video is already getting long. So,

I'm going to write out what didn't need

to be said in this video in the

description to finally actually get this

out. And this video has pretty much just

been B-roll. Anyways, anyways, thanks

again for watching. I'll see you on the

next