Some people have seen worse PB with Tesla Vision, some better, some no change. Many had no problems with radar and no problems after radar (like me). YMMV, so wait and see, and tell use what you find (good or bad).
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Phantom braking so bad I want to return my car
- Thread starter RHINO SKIN SOL
- Start date
I just received two updates virtually back to back on my non FSD vision only car. I was on 2021.36.5.5 for almost two months, and it still sucked both night and day on two lane rural highways... then I had 2021.44.6 for a whopping 4 whole days. It still sucked at night, and I only had one chance to drive it in daylight hours during that 4 days and it did seem to be improved over 36.5.5 with only 1 brake event in 40 miles of driving instead of 6 or more like I used to get... but one sample is not enough to be conclusive. Now (as of last night) I am on 2021.44.25.2
If it is as good as 44.6 during the day, I will consider it to be at least be "daytime useful"... but we will have to see how it is for night time driving. I suspect it will still suck at night, but it is good to see some improvement in daytime driving.
Keith
If it is as good as 44.6 during the day, I will consider it to be at least be "daytime useful"... but we will have to see how it is for night time driving. I suspect it will still suck at night, but it is good to see some improvement in daytime driving.
Keith
This may be counter intuitive to people, but in my experience in the non FSD vision only car is that the less distance it can see the better it drives (up to a point).
The more distance it try's sees the more freaked out it gets. At night it can see headlights coming towards you on a strait road for half a mile or more... but it has no idea how far away those headlight are, or if those headlights are coming in your lane, or in the opposing lane... so it freaks out. In a radar equipped car at night the cameras say "is that a threat?" and the radar replies "is WHAT a threat? I don't see anything." and the car drives on its merry way without hesitating.
I hope they get a handle on this, but judging distance without binocular vision is done by size comparison... during the day a car half a mile away is too small to register on the system as an object, let alone as a threat... but at night two point sources of light next to each other register on the system as an oncoming vehicle... but there is virtually no way to judge the distance to those point sources of light that are next to each other.
The solution will be seen as "unsafe" until they get tired of the system not working... the solution is to do it the same way a human does. When I see headlights a half a mile away I ignore them. I don't try to judge how far away they are, I don't try to figure out what lane they are in... I note that they exist and I ignore them until they are much closer.
There are three distance classifications at night.
#1 Way the F over there... not a threat, drive without a care in the world.
#2 Medium range... that car has not turned off the road since I first saw it and it is getting closer... I need to pay attention so I can assess the threat level when it gets closer (no fiddling with the screen, or looking at your phone).
#3 Close enough to judge distance and position. This is much closer than most people think and your reaction time with an oncoming car is miserably short if it turns out that they ARE in your lane.
As I said, the solution is to ignore oncoming headlights until they are close enough to judge distance and location... the "safety first" people will say that this is unsafe... but having a driving system reacting to something that it doesn't know the distance or location of as if it were a threat is unsafe as well.
Keith
sodaboy581
12/21 MY already wishing for Ryzen...
This is why I have zero confidence in AP or FSD on Tesla Vision only vehicles. You really do need that radar. I can't see a single camera, no binocular vision system, ever working properly, no matter how much AI you give it.
There's a reason we, as humans, have two eyes.
TheTheocracy
Member
Monocular vision utilizes the disparities between frames instead of instantaneous side-by-side frames from stereo vision. It is still possible to obtain 3D from a single camera, given that either the camera (car) or the background moves. It's a harder problem with the camera moving forward instead of sideways, but still can be done. There are several publications on depth estimation, 3D estimation, and the like for monocular vision. Stereo vision can provide better 3D rectification, but has limited distance as well, since spacing between cameras limits your ability to see those disparities, especially textureless things. I would argue that any able-bodied person can drive with one eye. How do people in 2D simulators drive (Forza, Gran Turismo)?
Vision only systems will probably rely more heavily on context than other modalities. What objects are there and their trajectories? Are some objects occluded? Without knowing the extent of what is being solved, we can't know what algorithms are being used to solve them. Even if we know the solution, training the system to mimic that solution is still a feat.
Especially with TACC, there might be much more context in the algorithm than a simple distance calculation. Estimate other potential objects coming in front from the side? Usage of other frame information? Is there a possible occlusion? RADAR information is only straight out in front with no context. I would venture to think that TACC already is capable of estimating distance similar to RADAR with less precision, but due to being 'traffic-aware' for incorporating more context, decision making is much more complex. Hence more false positives, but smarter and safer than specifically RADAR. But this would all be speculation.
Vision only systems will probably rely more heavily on context than other modalities. What objects are there and their trajectories? Are some objects occluded? Without knowing the extent of what is being solved, we can't know what algorithms are being used to solve them. Even if we know the solution, training the system to mimic that solution is still a feat.
Especially with TACC, there might be much more context in the algorithm than a simple distance calculation. Estimate other potential objects coming in front from the side? Usage of other frame information? Is there a possible occlusion? RADAR information is only straight out in front with no context. I would venture to think that TACC already is capable of estimating distance similar to RADAR with less precision, but due to being 'traffic-aware' for incorporating more context, decision making is much more complex. Hence more false positives, but smarter and safer than specifically RADAR. But this would all be speculation.
The more the better for sure, but where are these 5? Is it 3 in the rearview mirror mount and 1 in each door pillar? Can the 3 cameras in the mirror mount be used together to view the same object and do they have sufficient separation to make stereo vision effective? What about the pillar cameras? does their field of view overlap with part of the mirror mount cameras?
In other words "more than one camera" doesn't equal "binocular vision"
I was curious so I looked up some studies on binocular vision. Even humans don't get any significant advantage from binocular vision beyond 60 feet, and the binocular advantage is pretty small at 30 feet. Beyond 100 feet we rely completely on perspective, motion, relative size and other "context cues" to judge distance.
From what I understand FSD is using some aspects of binocular vision (when an object is behind another object from the perspective of one camera, but IS visible to another camera the processor sees the object rather than ignoring its existence) and for short range this is a huge advantage. The FSD tech is proprietary, so we as consumers will never know very much about how it works.
My point is that the camera only system will end up being wonderful in daylight conditions... but eyes don't work well at night and neither do visible light cameras.
They can either teach the system to "drive blind" at night just like humans do, not seeing or reacting too things that are beyond the reach of the headlights... or they can add other sensor systems (IR cameras, LIDAR, RADAR) for night time use. If they want to go visible light cameras only, then they have to accept the limitations of visible light.
Keith
PS: Just had a "perspective shift" of my own. Tesla doesn't see much difference between day and night because most of their testing is done in cities. The only time someone who lives in a city experiences darkness is if they are indoors in a room without windows (or with heavy curtains) with the lights turned off. People who live in cities don't really know what darkness is because they never experience true darkness when they are outside due to street lights and other light pollution.
Last edited:
TheTheocracy
Member
Here is a cool low light image enhancement video that shows you can still extract information from heavily compressed dark imagery. I can imagine Tesla not using it because it would be very computationally expensive. But maybe they can incorporate something into their feature extraction network?
If they have training data in the dark, their system may be able to pick up objects in low light, but having enough data to train and balancing the training data between day and night and all other scenarios would be tough. Some simpler algorithms, like Haar-cascades, work well in low light because it calculates a simple difference between areas, but the CNNs they use require lots of data and better training. Only if we can get explainable AI working well so the algorithms can tell us what they are seeing.
If they have training data in the dark, their system may be able to pick up objects in low light, but having enough data to train and balancing the training data between day and night and all other scenarios would be tough. Some simpler algorithms, like Haar-cascades, work well in low light because it calculates a simple difference between areas, but the CNNs they use require lots of data and better training. Only if we can get explainable AI working well so the algorithms can tell us what they are seeing.
While I am not saying Tesla vision can solve this, have you ever tried closing one eye and wandering around the room, or picking something up? Notice that you CAN do this pretty well even with one eye only? Binocular vision helps improve precision, but its by no means the only visual cue your brain uses to judge distance.
Binocular vision is essentially a precision tool for near-field and mid-field distance judgement .. that's why hunting animals, (dogs cats etc) have it, but prey animals (cows, horses etc) tend to favor all-round vision at the expense of binocular vision. Oh, and primates have it because its critical to jumping around in trees etc.
Very interesting, but dont forget cameras are better at seeing in low-light than our eyes anyway (yes, I know we have very sensitive eyes, but they can't handle the dynamic range, so e.g. a headlight can dazzle us).
OK so - since the 2021.44.25.2 update which I received on 12/23, I have noticed a change. I have been up and down the same 4 lane road in Central TX (I attached a picture of exactly where one of the typical PB incidents happens) - and so far - it has not recurred in 3 round trips. Now, this might be because there has been less traffic the last few days, or some other reason, but thus far I have hope that they might have figured out and at least partially corrected this problem? If anyone here has any similar experience, please let me know.
TheTheocracy
Member
Yeah definitely! I wonder if they have some sort of HDR, auto-brightness, or auto-ISO built-in. I'm not sure what kind of cameras Tesla use, but if image enhancement can give the algorithms any benefit or lessen the load of that learning the dynamic range within the object detection modules, it might be a possible avenue. I've done some object detection and person detection in low-light using traditional and deep learning approaches, and in most all cases, they benefit from some sort of enhancement, especially deep learning approaches.
Human vision has an excellent dynamic range but it's managed by the iris so response isn't instantaneous. Camera's can be blinded too. E.g. the repeater cameras at night given a turn signal or the pillar cameras whenever they face the sun. The specific behavior is a function of the hardware.
As a person with monocular vision I continuously surprise my spouse with what I manage but it's a significant operational defect. An example of this is knowing the distance to a closed overhead door -- my adaptations don't work well in the car. The Tesla can also make that determination. Oh wait ... it uses SONAR because using the right tool for the job is often a good idea.
Iain
Member
I live right across from your photo. Does your car handle the exit out of Sweetwater ever?
stopcrazypp
Well-Known Member
Tesla uses the AR0132 for most of its cameras. It does support in camera HDR, but that mode may result in a lower framerate and also introduces more latency and artifacts (as the chip needs to wait for 3 exposures and process it).
https://www.mouser.com/datasheet/2/308/AR0132AT-D-888230.pdf
There are newer sensors today that have a quad bayer array or a mode called DOL-HDR (which it switches exposures line by line) and can support single frame HDR (by pixels being set to different exposures) that avoids a lot of the problems.
In some sense that can be called dynamic range, but what is more important is instantaneous dynamic range, that is, the range of brightness that can be seen at the same time. If a camera (or eye) adjusts the amount of light entering that just moves the IDR higher or lower in brightness but doesn’t extend it.
Last edited:
Similar threads
