4K Cameras on new Model S?

[ucmndd]

You don't think they can put whatever lenses they want?

Alternatively bigger pixels would improve low light performance.

To what end? There is seriously zero reason to make these cameras 4K.

Nor is there any logical reason that the repeaters would have to be physically bigger to hold a 4K camera. The connection you’re trying to make isn’t there.

 

[mswlogo]

+1, this isn't a consumer TV, it doesn't need 4K for anything.

Right, that would be way to much data to process. It probably does have to do with the slightly bulging fenders.

I was wondering if they would move the cameras for better birds eye view mode.

 

[DanCar]

To what end? There is seriously zero reason to make these cameras 4K.

Wrong. What learnings or experience do you have in this field? Is two pixels better than 4? Higher resolution allows for vision understanding at greater distances, and thus allows for more time to react. There is a reason there is a vision exam for driver's license.

 

[ucmndd]

Wrong. What learnings or experience do you have in this field?

Enough to know that the computing power necessary to do real-time vision processing of potentially 8+ 4K video streams doesn’t currently fit in any compute envelope suitable for a car.

Enough to know that extra resolution/fidelity is likely not necessary for substantially better decision making given currently available software and computing power. There is currently no tangible autonomous driving benefit to being able to clearly make out the letters on the back of a trunk from 150 yards away.

There is also no reasonable expectation that advances in the near future will change this, particularly not to the point Tesla would waste a bunch of money “future proofing” their cars in this respect, especially given the aforementioned negative effects RE the noise trade off in high density sensors in low light.

Is two pixels better than 4?

Certainly not always, and basically never past a point of diminishing returns.

There is a reason there is a vision exam for driver’s license.

Yes, to ensure a suitable minimal level of vision to safely pilot a vehicle. There are no extra points for 20/10 or 20/15, nor is there any obvious correlation between better vision (again, past a minimally acceptable point) and safer driving.

 

[ucmndd]

It’s worth pointing out the signals from the current repeater cameras aren’t even 1080p… they’re 1280x960 at 36fps.

So the idea that they’d leap from that to something with ~6x the pixel density is even sillier.

There’s no “there” there.

 

[DanCar]

Enough to know that the computing power necessary to do real-time vision processing of potentially 8+ 4K video streams doesn’t currently fit in any compute envelope suitable for a car.

Any guesses what resolution Waymo cameras are? Hint: it is higher than 1080p.
Did you know that Elon said higher resolution cameras are coming with HW4?
Do you think it is smart for Tesla to put future camera hardware in current cars so that upgrade to HW4 is possible?

 

[ucmndd]

Any guessed what resolution Waymo cameras are? Hint: it is higher than 1080p.
Did you know that Elon said higher resolution cameras are coming with HW4?
Do you think it is smart for Tesla to put future camera hardware in current cars so that upgrade to HW4 is possible?

I’ve said my piece. By all means, speculate away.

 

[DanCar]

I’ve said my piece. By all means, speculate away.

There is no speculation in my assertions you just quoted. But speculation is better than incorrect assertions.

 

[S4WRXTTCS]

I'd love to get a better view of it to see what the optics looked like.

One of the frustrations I've had with the existing rear side cameras is the neural network has a tendency to put the cars in the wrong lane.

 

[Tigers]

IDK if 4k, but this is 100% HW4. During Autonomy Day in April 2019, Musk said they already started working on the 4th generation in late 2018/early 2018 and that they were roughly "halfway" through the development cycle. He said said it was "about two years away" which, yes, I know Elon time...however, that schedule puts us where we are now.

Combine that comment with the fact the computer has 10 TFLOPs of power (its going to need a badass CPU to run that stuff) and the larger and obviously modified cameras, it would make a ton of sense to roll out HW4 with these premium vehicles and yet another selling point for them to be worth $30k-40k more than the M3/MY LRs.

 

[stopcrazypp]

Any guesses what resolution Waymo cameras are? Hint: it is higher than 1080p.

The 5 cameras the previously Waymo used is 1920x1280, which is just slightly more than 1080p. It also uses pinhole lenses (which probably limits the resolution more than the sensor). Don't know what their 5th gen uses, but I doubt it's 4K (at least for the sides). 4K is useful for the front for more range, but I doubt the side cameras really need that much. It should be noted the camera array Waymo had does not even cover the repeater angle. There are three that covers the front, and two the sides (in similar position to the b-pillar cameras).
https://openaccess.thecvf.com/conte...riving_Waymo_Open_Dataset_CVPR_2020_paper.pdf

Did you know that Elon said higher resolution cameras are coming with HW4?
Do you think it is smart for Tesla to put future camera hardware in current cars so that upgrade to HW4 is possible?

I think it's smart to put better hardware, but I doubt that 4K would necessarily help. The module sticking out more probably helps more than the actual sensor resolution. But I guess we'll see soon. As a side point, I think for low speed parking lower angle cameras in the front probably would help too, but would require more changes to the bodywork.

 

[46&2]

Alternatively bigger pixels would improve low light performance.

No.

Not to dive too deep into photography tech, but, no. Low light performance is related to the absolute amount of light captured, not (mainly) to pixel size. Higher noise at pixel level evens out once scaled to the same resolution the other pixels, bigger to begin with, had. (Looking at the scene in different zoom ratios is NOT a reasonable way to compare image quality.)

 

[rooter]

No.

Not to dive too deep into photography tech, but, no. Low light performance is related to the absolute amount of light captured, not (mainly) to pixel size. Higher noise at pixel level evens out once scaled to the same resolution the other pixels, bigger to begin with, had. (Looking at the scene in different zoom ratios is NOT a reasonable way to compare image quality.)

No.

Can we start with the predicate that it is mandatory that the side cameras must perform extremely well in low light, if we are to have reliable FSD?

If so let's establish a few facts about enhancing low-light performance:
6 Benefits of Using Cameras With Larger Sensors

The new cameras can't be 4k as even if there were the processing power, it would consume resources which are better balanced to AI evaluation and decisionmaking. Even if they are 4k (which they're not), Tesla has more sense than to run them at that resolution.

The. cameras. are. larger. because. larger sensors have better low-light performance. Oh sure they'll have higher resolution, but the main point is larger sensors for better low-light performance.

 

[46&2]

A larger sensor is good for low light. Number of pixels is irrelevant (mostly). Nothing else said.

 

[DanCar]

A larger sensor is good for low light. Number of pixels is irrelevant (mostly). Nothing else said.

That is contrary to your earlier statements: ... not (mainly) to pixel size. And stating "No", when I said larger pixels improves low light performance.

 

[46&2]

I isn't, if you read carefully, again.

Bigger sensor = bigger area collecting ligh= better low light performance

More, therefore smaller, pixels on a given area = better capability of delivering higher image resolution

Less, and bigger individual pixels on the same sensor ares = less resolving power, but less data to process and so faster in time limited use cases.


The pixel size is indeed about irrelevant for low light scenarios. The sensor size is not.
The time of "bigger pixels = a lot better in low light" has been left behind long ago. Not speaking of the smallest tweaks, but of 'relevant' performance.

 

[DanCar]

The pixel size is indeed about irrelevant for low light scenarios. The sensor size is not.

You seem to have trouble with English. Larger pixels, does not mean more pixels. Pixel size is directly correlated with light gathering capability. The larger the pixel the more photons will be collected in a given amount of time. A larger sensor will not correlate with more light collecting ability if pixels have less light collecting ability, do to size of pixels or lens on pixel.

 

[127.0.0.0]

You would want a lower pixels and bigger sensor size to get good results when there is low light And large aperture but more like happy medium if say aperture is too large you gonna have a very narrow focus

 

[46&2]

You still don't read what I say or simply refuse to accept what modern camera technique has achieved. Nothing wrong with my english, here.

Pixel size is irrelevant for the low light capability of a sensor. Pixel density pixel size - both of course correlated on a given area - don't matter.

The "big pixel for low light rule" is dated. Wrong.

 

[46&2]

New You would want a lower pixels and bigger sensor size to get good results when there is low light

This is no longer true. Bigger sensor: yes, of course. Lower pixel count/bigger pixels: no.

Photographic Dynamic Range versus ISO Setting

Compare any 2 reasonably "same generation" sensors of a make you are familiar with. Many Companies sell high and low pixel count models for different use cases, which are usually separated by speed/saving data v.s. resolution capability/pixel count.

NOT the dynamic range and low light performance!