They're tiny lenses, like a cell phone camera, so the resolution is likely limited by diffraction in the lens. Adding more pixels won't help.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
AP2.0 Cameras: Capabilities and Limitations?
- Thread starter lunitiks
- Start date
Yet one arguable leader in consumer car automotive vision — MobilEye — is doing exactly that: adding more pixels. I would assume they have their reasons. To assume the amount of pixels eg in Teslas (be it AP1 or AP2) would be optimal would be arrogant.
Automotive lenses are not as small as cell phone camera lenses either. Yet for example Tesla’s lenses have ten times less pixels than the usual cell phone camera. It would seem like there is a middle ground where more pixels than eg Tesla has can help.
We should also not forget that MobilEye has a very efficient vision network. It would not surprise me if they can process a larger resolution with less power requirements than eg Tesla with their NNs. Tesla on the other hand has had to downscale even their inputs.
FWIW downscaled higher resolution input often is better quality than a sensor that produces the target resolution, depending on your downscaling algorithm.
Plus that way you always have access to the original resolution if you make a software breakthrough or want to bolt in out of band analysis on the frame that doesn’t need to run at high frame rates.
You can always downscale. But you can’t pull new pixels out of thin air. And Tesla is past the production volume where they can retrofit cameras.
Plus that way you always have access to the original resolution if you make a software breakthrough or want to bolt in out of band analysis on the frame that doesn’t need to run at high frame rates.
You can always downscale. But you can’t pull new pixels out of thin air. And Tesla is past the production volume where they can retrofit cameras.
Indeed, and for AI use you can run a simple edge enhancement filter before downsampling so that things like road signs are easier to read.
Does anyone know the specs on the original AP V1 system cameras that could read road signs? Or the current ones in the Niro/Kona that can read road signs?
At a cost in terms of low-light performance, of course. That's the caveat.
My best guess is that the manufacturer offered them a good bulk discount on the newer sensors. Also, bear in mind that cameras and other similar parts only get built for a certain period of time, and after that, the price per unit skyrockets to get the manufacturer to continue building legacy parts. So if you're going to design something that will be included in cars for several years, you always want to start with a current-generation camera, not something from a previous generation.
600 000 - 800 000 cameras a year is not an insignificant volume though. For that volume, it's easy to justify a seperate production line just for that type of camera. I doubt Tesla has taken anything less than 10 years into account when contracting the vendor.
I haven't looked to the numbers to confirm but AP2 was in production for about 9 months, so a rough guess would be 75,000 cars.
ap 2.0 and ap2.5 have identical sensor, just the color filter that's different. Also it's totally trivial to have support for both, Tesla just chose not to do it with a forceful assert in the code if the type does not match, but since they have identical NN so far - it's not really needed.
Yes. the physical color filters are different on otherwise same bit of silicon.
scottf200
Well-Known Member
[follow red arrow above to see thread/post where above came from]
Interesting that AP3.0 related cars and their new cameras may have a hard time with the red lights. From below and AP2.0 related cars (I think) you can clearly see the redish grey results on the red brake lights and red stoplights ... look a little orange.
In @DamianXVI's post you can actually see one of the windshield defroster threads (ref the fisheye-pic)
Similar threads
- Article
