HW4 will have next-gen cameras per Elon tonight.
No word on if it'll have more/different placement or not, and no word on retrofits yet
Yeah I will just keep my car forever until they retrofit out of spite
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Will Tesla need to retrofit side facing cameras?
- Thread starter not-Martin
- Start date
Yeah I will just keep my car forever until they retrofit out of spite
Well comparative pictures suggest they didn’t allow for visual obstructions
RobDickinson
Active Member
Again, how do humans cope when they dont have eyes on the front corner?
Atari2600
Active Member
I don't see how swapping out the side cameras would cost that much if they can use the existing harness for two cameras.
Daniel in SD
(supervised)
I guess they lean forward? I measured and the camera is 2 feet back from my face resting against the steering wheel. It really doesn't seem like that big a deal. The problems in FSD Beta right now are caused by the system not knowing when it can't see (which needs to be solved regardless of camera positions).
I think a bigger potential issue is the lack of a forward facing drivers side camera. It seems like that's necessary to see around cars in front of you.
They lean forward. You didn’t consider this?
And FSD is supposed to be BETTER than a human…
Yes. It does not know when it can't see. That's what humans do differently, we decide to adjust our eye position or car position. In some cases we creep forward to take a peek and quickly reverse. Sometimes we have to get out of the car and take a quick look or get someone else to look for us. If we really can't see and it's too risky we just don't go that way. I've done all those things at various times. FSD has been shown that it often just goes when it doesn't know it can't see. That's a strange logic: "it must be safe if I can't tell that it's unsafe".
stopcrazypp
Well-Known Member
They can do that probably because it's still L2. For L2 it's typically "better" to err on the side of less caution given people will be impatient if it hesitates too much (you have people overriding with accelerator). For L4 it would have to be tuned for caution though.
Didn't look at the video and pictures of verygreen's test previously, but that example precisely matches my expectations and why I said the angle of the car in various videos is likely necessary. If you tilt the car a bit to one side, the B-pillar camera would be in far better position to view around obstructions like that without actually intruding into the traffic lane. If instead you insist the car must be dead straight, then it's going to have to pull further forward to view around that corner.
For L2 it's typically "better" to err on the side of less caution
Ooh, that would be a tough statement for any manufacturer to make publicly. I'll disagree with that one
I'm not sure yet if FSD Beta is actually turning the car the wrong way to improve B-pillar views. Perhaps (as it might be a strategy that has some merit), but some of Chuck's drives show somewhat random behaviour at the exact same intersection. Certainly it's "creeping for visibility", I'll agree with FSD on that.
Daniel in SD
(supervised)
It's better to err on the side of less caution for YouTube videos. Either the car does something amazing and shows "confidence" or it fails spectacularly. And we know that everyone loves fail videos!For L2 it's typically "better" to err on the side of less caution
Ooh, that would be a tough statement for any manufacturer to make publicly. I'll disagree with that one
I'm not sure yet if FSD Beta is actually turning the car the wrong way to improve B-pillar views. Perhaps (as it might be a strategy that has some merit), but some of Chuck's drives show somewhat random behaviour at the exact same intersection. Certainly it's "creeping for visibility", I'll agree with FSD on that.
The argument I would make for less caution in an L2 system is that it keeps the user thinking "this thing is trying to kill me" which works better than any driver monitoring system. Of course I'm not convinced a system like FSD Beta can ever be safe* in the hands of the average driver.
*safe = the existence of system does not decrease overall road safety
stopcrazypp
Well-Known Member
Well you have people calling exactly for that in regards to ACC. They would rather have the car not react to as much things, than have phantom braking.For L2 it's typically "better" to err on the side of less caution
Ooh, that would be a tough statement for any manufacturer to make publicly. I'll disagree with that one
Yes, the behavior is a bit random, but there are plenty of cases where the car does turn fairly extreme angles and that does appear to improve visibility significantly for the cameras (I can probably do another diagram like I did for the other scenario in the other thread when I have the free time).
Sometimes, we play frogger and cause accidents.
Matching what humans do is already a difficult challenge. But in cases like this, we actually have an opportunity to do better.
As someone who has to creep to see around a neighborhood corner fence regularly, side-facing cameras in the corners (all four actually) could provide huge benefit. An AI system is only as good as its inputs.
Link? I only read new FSD computerHW4 will have next-gen cameras per Elon tonight.
No word on if it'll have more/different placement or not, and no word on retrofits yet
Or they could just not do anything and pocket the money. If pushed, they could just state they'd fix the software on the old cars to give FSD, eventually.
Knightshade
Well-Known Member
He specifically mentioned it along with HW4 during the AI day presentation
Tesla aims to launch Full Self-Driving Computer 2 with Cybertruck next year
Tesla plans to launch a new Full Self-Driving computer with Cybertruck electric pickup truck next year. The updated self-driving hardware...
electrek.co
No, they legally could not.
They couldn't do that unless they could provide evidence of active efforts to do so.
Which would be impossible if the problem is only solvable with hardware obviously
There are times where even doing right turn is a problem. To safely make right turn you still need to see left for a certain distance.
Need this in the front too. Need to see cross traffic when front camera views from side of vehicle are blocked by walls/fences/trees/parked cars without having to peep out 1/3 the length of the car.
The wide angle camera on front windshield is not enough because
1. the wide angle is too far back from the front tip of the car. Need camera locations at corner of car
2. the wide angle is not wide enough. Need 2 more wide angle cameras. One that can see directly left with 90-120 degree FOV and the other camera that can see directly right.
Yup but it even more than that human drivers also turn their head & eyes to the left 45 to 90 degrees so the left view is center of you FOV. This increases the accuracy of what you see. Analyzing image pixels at the wide angle front camera's FOV edge will be less accurate. Does the Tesla Vision pseudo radar accuracy drop when estimating distance of objects at the extreme edge of cameras FOV?
Last edited:
Yup, fender side cameras & b-pillar cameras need self cleaning
Now that Tesla neural network has pseudo-radar, radar or lidar are not needed.
Yes, more cams. Need 90 degree self cleaning side cameras at each front and back corner.
This is still a problem when trying to make right turns too when left view is partial blocked by blocked by walls/fences/trees/parked cars.
What common done
slow crawl of car into traffic lane with ability to reverse quicky
Severely angle the car to right more to get better view to left.
** unfortunately many never get good view and go into traffic lane anyway and then cause an accident.
Similar threads
