-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Rumor: HW4 can support up to 13 cameras
- Thread starter diplomat33
- Start date
CharleyBC
Active Member
Well, at least the repeater cam looks no better than my 4-year-old ones. The color is still poor; I assume your car is red, but it looks rusty orange drab, just like all red stuff does in 2018 repeater cams. And I don't think the resolution is any better.
Caveat: Software could be downsampling the video streams to conserve storage. That is, maybe the computer gets better resolution than we're able to see.
DanCar
Active Member
Zane Liu
Member
Haha yeah I have a nice rusty orange Model 3. Spray painted it myself
What does the video data in the raw file say about the resolution. That would answer the question.
Zane Liu
Member
stopcrazypp
Well-Known Member
The color is because of the color filter (RCCB) which Tesla is unlikely to change even if they bump up the resolution. The cameras are designed primarily for the computer, the repeater view on the screen and dashcam are secondary functions, so unlikely Tesla will change to favor that.
There’s always hope they change specs after a real hardware upgrade - for example the video became uncropped (slightly bigger field of view is what it felt like) and improved in quality when my previous 3 was upgraded from HW2.5 to HW3. But otherwise, like you said, that’s the same resolution and frame rate as the other cams
CharleyBC
Active Member
Sounds like something interesting to learn here. Please explain! How does this filter help the computer?
Knightshade
Well-Known Member
You might find this post informative
FSD & HW3 Clarification
I think that it’s a positive sign that EM is once again confirming that AP 2.0 and AP 2.5 owners who pre-purchased FSD will receive the AP HW3 CPU (Now the FSD CPU) upgrade at no charge. I am curious, however, as to whether or not us AP 2.0 owners will also require a Hardware upgrade of some...
teslamotorsclub.com
CharleyBC
Active Member
It is, to a degree. It doesn't explain what the advantage to the car is of such a filter.You might find this post informative
FSD & HW3 Clarification
I think that it’s a positive sign that EM is once again confirming that AP 2.0 and AP 2.5 owners who pre-purchased FSD will receive the AP HW3 CPU (Now the FSD CPU) upgrade at no charge. I am curious, however, as to whether or not us AP 2.0 owners will also require a Hardware upgrade of some...
Knightshade
Well-Known Member
You don't see the advantage of being able to more easily recognize the color green to a system trying to drive the car?
DanCar
Active Member
Can gather more photons compared to rgb. In other words, better low light visibility.
Are you asking about why any filter vs no filter? Without a filter, the camera is basically grayscale with the ability to measure how bright things are. Each filter tries to reduce the amount of light getting to the sensor to focus on certain colors like Red, so a red traffic light can be more visible to the sensor relative to a green light. So some sensors with a Red filter and others without ("Clear" filter) allow the combination to infer that a light more visible to Clear but not Red could mean it's a green traffic light. Because a filter prevent some light from reaching the sensors, it reduces their sensitivity in darker situations, but that's a tradeoff in being able to detect certain colors.
Here are some references in support of what @DanCar summarized.
Sensor | Imatest
Color filter array - Wikipedia
The meta-point here is there is a difference between “the best camera for the human eye” and “the best camera for an NN”. Just because a camera looks better on-screen doesnt necessarily make it better fort the NN. (And that is leaving aside that what the human sees has also gone through extra stages of processing just to get onto the LCD panel, with all its limitations.)
What you want from a camera for the NN is enough pixels so the camera can resolve everything it needs to and no more. Any more and the NN wastes compute cycles unnecessarily, limiting the speed with which is can handle tasks. You also want color resolution so that the car can unambiguously decode color-cued symbols, such as traffic signals, but (again) no more than that for the same reasons. It’s totally useless for the car to be able to tell 100 different shades of red that humans look for in a camera. I don’t pretend to know what the best camera specs are for ADAS systems, but I’d hazard a guess that no-one else does with any certainty since the entire field is rapidly evolving. However, “ohh that looks so much nicer” doesnt seem like a particularly scientific criteria to me
(I’m skipping over the secondary uses of the cameras for Dashcam et al where of course there are arguments for more human-like image presentation.)
I mean, people with red-green colorblindness can get driver's licenses (in most countries), which is why we have standardization of the position of traffic lights (red on top or left), at least in the U.S. I think that in countries where you drive on the left side of the road, the horizontal lights are in the reverse order, i.e. green is towards the oncoming traffic. Either way, it is consistent on a per-country basis, typically.
But RCCC does probably make it challenging (if not impossible) to tell the difference between a flashing yellow and flashing red light at night, when the bright areas are blown out, much as a colorblind person cannot do so even under the best of circumstances.
CharleyBC
Active Member
DanCar
Active Member
Similar threads
