Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register
This site may earn commission on affiliate links.
Those other 2 aren’t accessible in the app, do they look left and right or just straight ahead
IMG_4903.gif
 
I have posted this before, but to restate, they no longer use all 3 cameras with hw4 in the windshield. (The 3rd camera is a dummy, and does nothing) Hw3 does. They should in the same mount replace the wide angle, or near camera, or both with 1 or 2 dedicated cameras looking left and perhaps right.
This would leave 1 forward looking camera that would be complemented with the right and left views. No issues keeping them clean, and both hw3 and hw4 could easily be retrofitted with this. The cabling and connections exist today. Look at the wide angle and near angle in park. They are basically useless. DM me Elon for the invoice.
Agree 100%. I've also suggested this. Even if you have to modify the window area it would be a significant improvement and add safety to FSD.
 
This looks more like a PR boosting feature than a practical feature. There are sooooo many exceptions and limitations that I suspect it will be more of a "special operation" feature than a usable feature. In fact IMO it may be more of a hindrance feature that a practical feature. If using in daylight without sunglasses it will probably work most (but not all) of the time. For me it needs to be cloudy or hazy to not wear sunglasses in the day. At night or with sunglasses it will probably be hit or miss and a LOT of MISS. So likely it may go some with no nags and then unexpectedly it will loose track of your eyes and nag you. This may be more trouble than just continue with the expected nags we have now. To me it either needs to be NO nags like Ford and GM or just leave it like it is since it is not a proper eye tracking system..
Unless non-polarized sun glasses will work here is an alternative.
Tesla sun visor.jpg
 
What is controlling for between the lines overall or primarily?
The neural network processes all the pixels from all the cameras in order to figure out how to do anything. Which pixels have the greatest role to play for which tasks can only be known by tearing apart the neural network and watching which neurons are firing when performing each task. In truth, there's no way of knowing how the neural network keeps a lane. It may rely on the vehicle in front of it for much of its lane keeping. Or it may use its distance from the side of the road. It isn't trained to approach the driving task the way we do. It's only trained to duplicate a certain style of driving, and the neural network weights produce that style. Or so goes the theory.

Sorry, what does that mean?
The "repeaters" are the front fender cameras. The blind spot cameras. I don't know where the name comes from, but here's a TMC thread on the topic.

 
The neural network processes all the pixels from all the cameras in order to figure out how to do anything. Which pixels have the greatest role to play for which tasks can only be known by tearing apart the neural network and watching which neurons are firing when performing each task. In truth, there's no way of knowing how the neural network keeps a lane. It may rely on the vehicle in front of it for much of its lane keeping. Or it may use its distance from the side of the road. It isn't trained to approach the driving task the way we do. It's only trained to duplicate a certain style of driving, and the neural network weights produce that style. Or so goes the theory.


The "repeaters" are the front fender cameras. The blind spot cameras. I don't know where the name comes from, but here's a TMC thread on the topic.

thanks, so when I look at my “repeaters” the image of the side of the car that it sees is NOT uniform in the REAR+REPEATERS camera view. Could this cause a problem then with in the lane positioning issues? Other than in car SW “calibration” is there some other control for this ? I would think that IF the car sees that the image of the side of the vehicle represented is not same same on both sides, it would either throw and error or require corrective action, OR make some calibration adjustment… and yet, this doesn’t seem to help at all with out of center of lane proper bias.
 
I would think that IF the car sees that the image of the side of the vehicle represented is not same same on both sides, it would either throw and error or require corrective action, OR make some calibration adjustment… and yet, this doesn’t seem to help at all with out of center of lane proper bias.
Are you seeing lane centering issues on interstate highways? My highway centering is always rock solid, but on secondary roads, the car wanders. That different is down to V11's heuristic approach versus V12's neural network approach.
 
Are you seeing lane centering issues on interstate highways? My highway centering is always rock solid, but on secondary roads, the car wanders. That different is down to V11's heuristic approach versus V12's neural network approach.
Main highways and even more so local roads with markings, yellow center and white right - usually a bike lane. but highway too, just slightly less ~ 4” in that case and not 6 or more likely 8-10” off.
 
Have you tried turning FSD off on the Autopilot screen? My wife's profile has FSD off, and she gets TACC, double pull for lane keeping. I think there are visualization differences too. So, you may be able to revert from locked out FSD to plain old AutoPilot.

I have no strikes, and my SO is out with the car, so I can't verify this suggestion, sorry.
The FSD button was automatically disabled when I got 5 strikes. Both Autopilot and FSD buttons are disabled now for the week.

It is not fair, that not only Tesla took one week off of FSD usage ($25 worth) from my monthly subscription, they even disabled Autopilot usage too. :mad:
 
Main highways and even more so local roads with markings, yellow center and white right - usually a bike lane. but highway too, just slightly less ~ 4” in that case and not 6 or more likely 8-10” off.
I'm sure the system can tolerate some variation in the camera orientation, but if it's too far off for even a camera recalibration to sort it out you're probably best off taking it in to have service look at it. I doubt that they'd check the FSD performance, but they can check to see if the cameras are physically out of spec.
 
This thread was just a bit tired with everyone waiting for v12.4 videos to drop. Aren't you glad I brought up the B-pillar camera issue so we'd have something different to discuss :). In a few days (hopefully) it will be v12.4 all the time!

The only thing better would be talking about why it takes so long to release an update. Maybe it's a never ending need to add simulated training data?

What FSD features will Tesla need to punt on and replace with simulated training data? I'll bet they've been working on Chuck's unprotected left.

I suspect v12.3.6 currently uses more simulated training data than TSLA would ever want to admit like acceleration, deceleration, stop sign intersections, .... Maybe anything that's unrefined and clunky feeling.
 
thanks, so when I look at my “repeaters” the image of the side of the car that it sees is NOT uniform in the REAR+REPEATERS camera view. Could this cause a problem then with in the lane positioning issues? Other than in car SW “calibration” is there some other control for this ? I would think that IF the car sees that the image of the side of the vehicle represented is not same same on both sides, it would either throw and error or require corrective action, OR make some calibration adjustment… and yet, this doesn’t seem to help at all with out of center of lane proper bias.
See the difference in the image here, left vs. right side and the lane markings?

Clearly the RIGHT camera is seeing about 50% More car side than the left repeater?

As well, the don’t seem to be seeing the same area, at least not from the bottom or top to the hip line. This shouldn’t happen.

IMG_0122.jpeg
 
Sorry, what does that mean?
1716845613866.png

All newer Tesla's have one of these in each side. They are the primary lane centering cameras used. They do use others to assist, but these are the main ones.
Using the car in front of you, or a follow me car doesn't always work if they are drunk for example and not keeping in their lane. The repeaters work excellent in little or no light situations because your taillights give them enough light to see the lane markings. Shout out to the bad ass Tmc member who found this image a few pages back. If you are having any lane centering issues that a recalibration doesn't fix, service center is the way to go. V12 loves to wobble a little during city driving.
Highway stack you should be seeing almost no wobble.