Estimated Tesla Autopilot miles reaches 1.3 billion

[Tezlanian]

Amon Shashua has said multiple times that it has. if you actually watched Mobileye presentation from 2014 till now you would know they have since stopped talking about sensing. Its of no interest to them anymore. They have basically solved it for SDC. Nowadays for quite a while they only give talks on their RSS (Responsibility-Sensitive Safety).

First of all the question of " solving vision or solving perception" as most people put it is wrong. you don't need to solve vision to get SDC.
If you were trying to create a AGI then yeah. For a SDC you simply need a system that can reach a certain level of accuracy/verification.

for example the mobileye's eyeq3 1 false positive of pedestrian detection every 400,000 miles according to Amon.

in reference to sensing aka solving vision for SDC, amon said
"When people think about sensing, they think about object detection, vehicles, pedestrians, traffic signs, lights, objects, etc. You receive an image as an input and your output is a bounding box...this is the easiest problem. this problem has been solved."

In reference to second level of sensing (semantic free space) where you can drive and where you can't drive:
"this is already in production.

in reference to the third level of sensing (drive-able path) where your input is an image and the output is a story. For example this lane is ending in 15 meters, etc.

Amon says its an open problem in the industry that is solved by REM maps.

So yes vision for SDC is SOLVED.

To see how ahead Mobileye is in respects to sensing. Check out for example the bounding box accurary of Zoox's sensing system.
Notice how inaccurate and jumpy the detection is?

You can do this by simply going Zoox website. They have a video on their home page

https://zoox.com/wp-content/uploads/2018/07/Vision_Video_graded-Window-3.mp4

more loose boundary box

Now compare that to the tight and accurate 3d bounding box of mobileye eyeq4.
The accuracy between the two is night and day. Its not even comparable.

"Perception" in the most simple sense you're trying to use was already solved over a decade ago with algorithms like Speeded up robust features - Wikipedia among others. I think we're talking about different things. All of these systems are trained or utilize a discrete algorithm to recognize an object or use machine learning. If they have not been trained on the data or not deliberately instructed to recognize an object then it won't recognize it. That's how the Harry Potter watching guy got decapitated while Mobileye's system hummed along a-OK.

Recognition and perception are different things. Perception is an actual understanding of that environment and the machine only knows what it has been told. It will only work in the scenarios it was designed for. For anyone to claim it is completely solved is pretty asinine. Perhaps for a very limit environment, however, that's all. Intent of vehicles in round abouts, ice over traffic lights, traffic lights being out of power, a plane emergency lands on the interstate, etc. I highly doubt all this is included in their wholly solved perception model. Solving most of your problem space while ignoring everything else doesn't mean it's solved.

 

[Tezlanian]

This is pretty neat, though...

Guy in the back seat looks like he is playing a videogame and watching his monitor a lot... driving the car with a controller? Demand to see his hands!

 

[Bladerskb]

Wow where are these photos and info coming from about this BMW?

I added the info. although after further research i'm not so sure on the lidar.

You are saying every single 2019 BMW car is going to have an EyeQ4? How do you know this?

Well i shouldn't say 2019 BMW (since the germans use diff year for naming) but more "BMW produced in 2019". The models i have checked all have tri-focal camera and driver facing camera. Some "2019" models released this year might use eyeq4 like the 2019 X4. But they use single camera config. I have yet to see a model released so far this year with tri-focal camera.

2019 BMW 3 series (G20)

https://i.imgur.com/yksVp30.jpg
1m26s

2019 BMW 8-Series M850i gran coupe
https://car-images.bauersecure.com/...series_052.jpg?mode=max&quality=90&scale=down
https://images.hgmsites.net/lrg/2019-bmw-8-series_100654606_l.jpg


2019 BMW 3 Series Wagon

 

[NateB]

In addition to the training data discussion, I am very interested in this counter for another reason: safety statistics. tldr; It looks to me like autopilot fatality rate so far is quite good, though it is hard to say exactly how good.

From Wikipedia

The National Safety Council, using methodology that differs from the NHTSA, reports a rate (including deaths of pedestrians and cyclists killed in motor vehicle accidents) of 1.25 deaths per 100 million vehicle miles (or 12.5 deaths per billion vehicle miles) traveled in 2016.
​

I'm sure that this is not a perfect apples-to-apples comparison, but if an average autopilot mile was just as safe as an average 2016 US mile in another vehicle (and average occupancy was the same), we would expect approximately 16 deaths (including pedestrians and occupants) after 1.3 billion miles driven. How many autopilot deaths have been reported so far? 3 or 4? It seems to me that autopilot as implemented so far, despite all of its flaws, including some driver complacency, is no more dangerous than the average human driver. On the contrary, it might be significantly safer, although to confirm this we would need to control for average occupancy in Tesla vs. other vehicles, differences in fatality rates on the average autopilot mile vs. the average US vehicle mile, and differences in safety of Teslas vs. other cars.

 

[Bladerskb]

@Bladerskb -- I've seen this show before with the Audi "amazing world's first L3" using MobilEye's latest and greatest. Wake me when it ships and actually does what it says. Its all smoke and mirrors. BMW's lane keeping right now is COMPLETE GARBAGE. EyeQ5 couldn't fix that hot mess.

Well Audi A8 uses eyeq3 and only one front facing camera so it was definitely not the latest and greatest.
Secondly L3 is NOT legal in the US. However there are L3 cars using mobileye planned for 2019 in Japan, Germany.

There are no smoke and mirrors. While 2017 BMW lane keeping might not be impressive to you. You haven't seen 2018 let alone whats shipping in 2019 with eyeq4 and trifocal camera (the fact that it looks like they are going the supercruise route and doing full hands free means they have finally allocated the sufficient number of engineers to work on their ADAS.)

EyeQ4 is a level 5 capable chip. But the end result is always determined on how much resource you devote to developing with it.

Most auto companies have a few engineers working on their ADAS system while they have hundreds working on their SDC prototype (most being external). So that's why you have such polar difference between what a company has as their SDC vs their actual ADAS. You have garbage systems using eyeq3 from the likes of propilot, honda, toyota, ford and then a vastly better system in gm supercruise and AP1.

You have few companies who actually allocated sufficient resource to ADAS. Companies like GM, Volvo and ofcourse Tesla.

Tesla being unique because their entire SDC team IS their ADAS team. So you had hundreds of people working on level 2 ADAS which was unprecedented. But the caveat being, while other companies ADAS isn't representative of their SDC progress. Tesla's ADAS is their SDC tech.

 

[Bladerskb]

Demoware is demoware. I wonder if raw data is possible to get out of those boxes to doublecheck the results.

But the eyeq4 isn't demoware though, seeing its in production cars today.

You mean vector data from the EyeQ chips or the actual raw image input?

I mean I agree they send less than 1% of camera data, but 1% seems to be super generous since it's probably even less than 0.1% too (unless we add a lot of other qualifiers)

Funny enough i used to say 0.1%(which is closer to the actual percentage as you said), but people get too offended about the cold hard truth. So i now say less than 1%. Which is alil easier to swallow.

But does it even matter? if People like Fred from Electrek, @Trent Eady and @Lex_MIT won't amend they publications?

The counter from @Lex_MIT leads to articles like these being written. Tesla shareholders who sell if EV maker goes private will be losers for one big reason: A.I.

Tesla is the only company truly advancing AI in its vehicles. Musk made the decision to invest in real-world data analytics early. This leap in technological advancement has left competitors like Alphabet's Waymo, Uber, BMW and GM scrambling in vein to keep pace.

To understand how Tesla is unique with AI, you need to understand that Tesla has been collecting billions of miles worth of real-world driving data since 2012, through Model S, X and 3 drivers.

Billions of new sensors translate into yottabytes of real-world road data, all of which is sent directly to the cloud. Crowdsourcing, or "the wisdom of the crowd," provides Tesla with minute details.

Musk knows his cars will be fully autonomous, fully thinking by 2019, while his competitors are still struggling with low-competency software and expensive, unreliable lidar systems. All Tesla vehicles since 2012 (S, X, 3) were built with the potential to one day become self-driving.

Fully thinking, Fully sentient by 2019. lol

What do you think about this @verygreen ? Mostly the fact that people like Fred from Eletrek posts articles based on your findings alot but seems to ignore what the overall result means.

 

[strangecosmos]

In his talk at CES 2018, Amnon Shashua said that semantic segmentation for lanes and driveable paths is “largely an open problem from a perception point of view.” So, not quite solved yet!

While 2017 BMW lane keeping might not be impressive to you.

Do 2017 BMWs use Mobileye tech? The IIHS tested a 2017 BMW 5 Series and found that it was incapable of doing lane keeping on hills. It stayed within the lane line 0% of the time.

The Model 3, by contrast, only had one error where it touched the line. It succeeded on 17 out of 18 attempts.

 

[Bladerskb]

Since all he did was post a counter that is intended to give an idea of how much AutoPilot is used, and since his research is focused on human interaction with driver-assistance systems, I fail to see the relevance of the data upload to the counter.

"the data streams in these vehicles (whether under manual or Autopilot control) is available to be used for training the neural networks that perform the various components of the perception-control task."

Which leads to articles like these:

"Tesla is the only company truly advancing AI in its vehicles. Musk made the decision to invest in real-world data analytics early. This leap in technological advancement has left competitors like Alphabet's Waymo, Uber, BMW and GM scrambling in vein to keep pace. To understand how Tesla is unique with AI, you need to understand that Tesla has been collecting billions of miles worth of real-world driving data since 2012, through Model S, X and 3 drivers. "

"Billions of new sensors translate into yottabytes of real-world road data, all of which is sent directly to the cloud. Crowdsourcing, or "the wisdom of the crowd," provides Tesla with minute details "

"Musk believes it will take about 6 billion real-world miles to gain "worldwide regulatory approval" of true self-driving technology. Tesla is the only company to already have surpassed that mark. "


Yottabytes of real-world road data!

In his talk at CES 2018, Amnon Shashua said that semantic segmentation for lanes and driveable paths is “largely an open problem from a perception point of view.” So, not quite solved yet!



Do 2017 BMWs use Mobileye tech? The IIHS tested a 2017 BMW 5 Series and found that it was incapable of doing lane keeping on hills. It stayed within the lane line 0% of the time.

The Model 3, by contrast, only had one error where it touched the line. It succeeded on 17 out of 18 attempts.

Firstly, He said numerous times that its a open problem for the industry. Whenever he says something is an open problem. He's referring to the industry not to mobileye. Hence "open". The steps they have taken to solve it is their REM Maps. Driveable path is apart of Rem maps solution.

Furthermore the sensing you refer to and most people is object detection and classification which has already been solved, in-addition to free-space. Driveable path is used for planning and retrieved by the path other cars before you have taken, there are other ways to retrieve it using HPP which is also in eyeq4, and more driveable path extracting algorithm. They also have network that tags semantic lanes (which lane leads where and what light/sign belongs to what lane). The fact that they have a map gives them the ability to make manual updates if they so desired.

"RoadBook™ provides a highly-accurate, rapidly-refreshed representation of the static driving environment. This includes road geometry (lanes, drivable path, paths through complex intersections), static scene semantics (traffic signs, the relevance of traffic lights to particular lanes, on-road markings), and speed information (i.e. how should average vehicle speed adjust for curves, highway ramps, etc)."

Rem Mapping is the solution to driveable path.


On the other hand Supercruise uses mobileye and its the best L2 hands down. its not even close and this is from 5 independent reviewers which compared it to Tesla Autopilot.

Companies devote different resources to their ADAS team which is why honda, toyota, ford systems sucks and why supercruise is transcendent.

"Perception" in the most simple sense you're trying to use was already solved over a decade ago with algorithms like Speeded up robust features - Wikipedia among others.

I think you are failing to realize what sensing for SDC and the accuracy required. We are talking 99.9999%+ in hundreds of categories, including scene understanding, and intent. We are not just talking about a simple bounding box. But a 3d environmental model, including dimension, orienttion, speed, velocity, distance, state of every object. Blinkers on? Door open? etc...

All of these systems are trained or utilize a discrete algorithm to recognize an object or use machine learning. If they have not been trained on the data or not deliberately instructed to recognize an object then it won't recognize it. That's how the Harry Potter watching guy got decapitated while Mobileye's system hummed along a-OK.

The eyeq3 doesn't have side detection. second of all, there are other networks that handle that such as the semantic free space network.

Notice how the SFS network green carpet doesn't cover the car? That's because the network has been trained to detect roads. The side of a trucker trailer is clearly not a road. But SFS weren't being used by tesla to stop the car, only to drive where lanes are abiguous which makes sense with it being a ADAS. But in SDC, SFS is heavily relied upon.

Recognition and perception are different things. Perception is an actual understanding of that environment and the machine only knows what it has been told. It will only work in the scenarios it was designed for. For anyone to claim it is completely solved is pretty asinine.

Hate to brake it to you but it is solved.

Perhaps for a very limit environment, however, that's all. Intent of vehicles in round abouts

Intent of vehicles is territory of planning (driving policy) you should watch Mobileye presentation on the double lane merge.

, ice over traffic lights, traffic lights being out of power,

Again that's planning not sensing. Cruise car already navigate non functioning traffic lights.

a plane emergency lands on the interstate, etc. I highly doubt all this is included in their wholly solved perception model. Solving most of your problem space while ignoring everything else doesn't mean it's solved.

SFS easily sees that there is a object in the way, a pretty big one and Lidar easily sees it.

 

[verygreen]

But the eyeq4 isn't demoware though, seeing its in production cars today.

You mean vector data from the EyeQ chips or the actual raw image input?

Both. Vector data and raw images so people could do their own visualization to verify the accuracy.
It's all too easy to produce nice demoware (see the FSD video from Tesla in 2016).

Fully thinking, Fully sentient by 2019. lol

What do you think about this @verygreen ? Mostly the fact that people like Fred from Eletrek posts articles based on your findings alot but seems to ignore what the overall result means.

Strange article, "All Tesla vehicles since 2012 (S, X, 3) were built with the potential to one day become self-driving" huh? I mean I know there are some ap0->ap1 retrofits (and some comma.ai work lately), but it's still quite a stretch to claim that. Also Tesla does not offer any official retrofits either.

Anyway, I don't think we even understand what sentience is, so how can we have fully sentient machines next year? But then again, I am no AI/NN expert so I don't know what I am talking about here.

Do 2017 BMWs use Mobileye tech? The IIHS tested a 2017 BMW 5 Series and found that it was incapable of doing lane keeping on hills. It stayed within the lane line 0% of the time.

Well, I had an AP1 loaner for some time, it's unusable on hilly roads, so I guess that checks out if they use MobilEye 3 or whatever it was.
Also reportedly Tesla really pushed he limits on their units so other manufacturers might perform even worse (and I can easily believe that too).

 

[strangecosmos]

for example the mobileye's eyeq3 1 false positive of pedestrian detection every 400,000 miles according to Amon.

What about false negatives? That’s what I’m more worried about as a pedestrian!

An error rate of less than once per 1.3 million miles seems like a good first target, since that’s the average rate of injuries and death caused by human drivers. Make it 2.6 million miles to beat humans 2x, or 13 million miles to beat us 10x.

Not all errors cause collisions, either. So this would be a conservative way to benchmark errors.

However, fatalities happen once per 91.7 million miles, so that would be the ultimate benchmark. You would want to make sure your autonomous system doesn’t cause more fatalities even while causing fewer injuries.

Firstly, He said numerous times that its a open problem for the industry. Whenever he says something is an open problem. He's referring to the industry not to mobileye.

In the video, he is describing Mobileye’s technology. He explicitly mentions EyeQ4 and REM.

 

[Tezlanian]

Hate to brake it to you but it is solved.

Again that's planning not sensing. Cruise car already navigate non functioning traffic lights.

SFS easily sees that there is a object in the way, a pretty big one and Lidar easily sees it.

I doubt novel objects are dealt with well or at all in their tech, meaning perception is far from solved. Perhaps they could demo a giraffe grossing the street instead of a person. A plane landing in front on a long stretch of road, etc. Without tech in the wild to be independently verified I doubt all this very much. I am sure the tech works fine within the confines of what they consider perception worthy but it's far from solved.

And those situations are not acting up information, that's understanding the environment in the first place to act upon. Let's see what that plane landing in front of the car actually looks like to their system. This is a realistic scenario for me as it's happened twice in the past 8 years right near where I live.

 

[Bladerskb]

What about false negatives? That’s what I’m more worried about as a pedestrian!

An error rate of less than once per 1.3 million miles seems like a good first target, since that’s the average rate of injuries and death caused by human drivers. Make it 2.6 million miles to beat humans 2x, or 13 million miles to beat us 10x.

Not all errors cause collisions, either. So this would be a conservative way to benchmark errors.

However, fatalities happen once per 91.7 million miles, so that would be the ultimate benchmark. You would want to make sure your autonomous system doesn’t cause more fatalities even while causing fewer injuries.

Mobileye has a false negative of near zero for pedestrian detection and that was from years ago according to amon (CVPR 2016, launched in 2010). I don't know the actual percent, they haven't released it yet, i'm sure they will sometime next year to prove their RSS works.

Amon showed humans the errors, but even humans had it wrong. That was pre-deep learning by the way but today Mobileye's pedestrain detection is even better because of context that deep learning network provides.

For your number, you have to keep in mind that in the accidents that humans hit pedestrian are almost always easy cases for detection networks and almost always due to human distracted driving.

But since Amon already said that sensing is solved. They should be able to beat any benchmark. Remember that sensing can be benmarked offline. You don't need to be driving autonomously on the road to do that.

"Mistakes of the sensing system are easier to validate, since sensing can be independent of the vehicle actions, and therefore we can validate the probability of a severe sensing error using offline data,"

Also noticed how Amon differentiated error vs severe sensing error. Not detecting a pedestrian hanging on the side of a 4 story building is different from not detecting a pedestrian on the road or about to enter the road.

In the video, he is describing Mobileye’s technology. He explicitly mentions EyeQ4 and REM.

Driveable Path is not sensing. Its under the category of mapping. All companies use HD map and don't require live parsing of what traffic light belong to what lane and what lane leads to what, what lanes are complex versus what lanes are simple.

Driveable path can be retrieved by the cars that have already gone through that same road before and the path they took (ex: angle/radius of turn).

Driveable path is not apart of sensing environmental model, but mostly used to create automatic HD maps , redundancy for maps and for driving if you don't want to use maps at the moment.

Its an open problem to the industry and as he highlighted. its an open problem because no one else does it. Everyone else just uses HD maps. Its not an open problem to mobileye though because they have been working on it for quite a while. But because every road around the world and even in the us is so different. I don't think you will have a network that can adequately tag them. Highways are easy but urban roads are something else. This is why the HD map only solution will always be the dominate solution going forward for another decade. You can still parse simple lanes and road edges, curbs, barriers, stuff that mobileye already solved.

Mobileye has been working on it for quite a while, its definitely not an open problem to them.

Besides, when talking about sensing, perception, building a 3d environmental model. Amon says "When people think about sensing, they think about object detection, vehicles, pedestrians, traffic signs, lights, objects, etc. You receive an image as an input and your output is a bounding box...this is the easiest problem. this problem has been solved."


Whenever Amon talks about sensing mistakes, he's not talking about driveable path, angle of turn, etc. He's says "missed a vehicle, pedestrain, or had wrong measurement of them"


Sensing is SOLVED. Driveable path is not sensing although it uses sensing algorithms among other things but its not apart of the environmental model.

 

[ItsNotAboutTheMoney]

"the data streams in these vehicles (whether under manual or Autopilot control) is available to be used for training the neural networks that perform the various components of the perception-control task."

You're making the leap from "is available" to "is being used". His research group isn't even focused on autonomy. It's focused on human-AI interaction, so the page actually more useful information about the frequency of use of Autopilot.

But of course immediately Trent made the same leap and now the thread has degenerated into the usual pissing contest on autonomous vehicle progress where people get overexcited about big numbers that when translated into reality still imply that any unsupervised AI-driven vehicle is still too dangerous to allow on the roads.

Anyone want to discuss the implications of increased use of driver assistance systems, and the relative merits of various approaches of restriction and monitoring driver attention?

 

[Snuffysasa]

I added the info. although after further research i'm not so sure on the lidar.




Well i shouldn't say 2019 BMW (since the germans use diff year for naming) but more "BMW produced in 2019". The models i have checked all have tri-focal camera and driver facing camera. Some "2019" models released this year might use eyeq4 like the 2019 X4. But they use single camera config. I have yet to see a model released so far this year with tri-focal camera.

2019 BMW 3 series (G20)

https://i.imgur.com/yksVp30.jpg
1m26s

2019 BMW 8-Series M850i gran coupe
https://car-images.bauersecure.com/...series_052.jpg?mode=max&quality=90&scale=down
https://images.hgmsites.net/lrg/2019-bmw-8-series_100654606_l.jpg


2019 BMW 3 Series Wagon

Hmm, thanks for this info.

I doubt the Lidar but that sensor does look funny.

Good catch about the 2019 bmws having trifocals, defs eyeQ4.

Driver monitoring doesn’t mean much though, driver monitoring could be in an eyeQ3 system or no eyeQ at all system.


Also skimming through this thread...
Have patience,

Soon all the Tesla fans will realize how far behind they are from Mobileye’s upcoming systems.

 

[strangecosmos]

Lex teaches an MIT course on self-driving cars, publishes research on self-driving cars (not just ADAS), and is working with others at MIT on developing a self-driving car prototype. He’s an expert on the subject, so I was hoping he might weigh in on the question of how much Tesla’s Hardware 2 fleet might accelerate its development of full self-driving.

Lex did say, “Tesla Autopilot has driven over 1.2 billion estimated miles. This is a remarkable accomplishment in the history of applied AI research, development, and deployment.” But I was hoping he could expand on that. It is a rare privilege to hear an expert’s opinion.

Andrej Karpathy’s recent talk suggests that collecting and labelling camera data is a core part of Tesla’s AI strategy. Karpathy didn’t discuss this explicitly, but video data is useful not just for object recognition but also for path planning. If you collect and label video of vehicles, bikes, pedestrians, etc., you can train neural networks that do prediction. Prediction is an important part of path planning. Knowing a nearby car’s likely trajectory lets you safely plan a path that doesn’t intersect with that car’s likely path at any future time step.

Labelled camera data is also useful for semantic segmentation, which is also important for path planning. It lets the car know where in space the driveable road is, as opposed to sidewalk, construction cones, cars, bikes, pedestrians, etc.

I imagine radar data is useful too, but I don’t really know anything about how exactly it might be used to train neural networks. I’m super interested in learn more if anyone knows anything about that!

We have only small glimpses of information about what data Tesla collects, and how, and why. We know that Tesla has triggers, e.g. a trigger that uploads sensor data whenever a construction zone is detected. It could also do purely random triggers, or triggers that go off whenever there is an unrecognized object, or an object the ConvNet has low confidence about.

As Karpathy mentions in his talk, it would not be helpful to collect all the data that passes through Teslas’ sensors. It would mostly be repetitive recordings of highway driving, since that is what most driving is. This is just a feature of the problem space; it is not unique to Tesla. Astro Teller at Google X talked about how Waymo is always hungry to find new stuff it hasn’t encountered. That’s a big reason for structured testing (what happens at Castle) and simulation.

The main practical difference between Waymo and Tesla in this regard is that Waymo records everything and most of it presumably isn’t helpful, and Tesla records a tiny fraction and tries to make sure that tiny fraction is helpful. So aggregate miles driven is a fair comparison...

...if Tesla is actually effective at recording what’s helpful and not missing important data. A big if. The triggers have to be good at capturing everything new and interesting.

 

[verygreen]

We have only small glimpses of information about what data Tesla collects

Not really all that small. We have a decent window into this whole thing.

triggers that go off whenever there is an unrecognized object

I am not even sure this is possible. How do you (system) even know there is an object if it was not recognized?

object the ConvNet has low confidence about.

While remotely possible, I don't think it works like this.

The main practical difference between Waymo and Tesla in this regard is that Waymo records everything and most of it (presumably) isn’t helpful, and Tesla records a tiny fraction and tries to make sure that tiny fraction is helpful. So aggregate miles driven is a fair comparison...

The problem is you don't know what's helpful and what's not and things that are not helpful NOW might become helpful later.

...if Tesla is actually effective at recording what’s helpful and not missing important data. A big if. The triggers have to be good at capturing everything new and interesting.

Triggers are by definition NOT capturing anything new. They are mostly reactive. "oh, how does our lane detection work on curves? let's capture some data when steerign angle is above X and speed is above Y", or "We have a new NN label 'zombie', lets add a trigger that captures first 2 occurrences of this label from a bunch of cars, and then manually sift through those to see if there are any false positives" (cannot even do false negatives with this approach!)
The closest to "new" it gets is the intervention triggers - "the car is on autopilot and XXX and the driver applies brakes/steering/acceleration"

 

[strangecosmos]

Not really all that small. We have a decent window into this whole thing.

Really? What did I miss? I haven’t been keeping up with all the threads.

Triggers are by definition NOT capturing anything new. They are mostly reactive. "oh, how does our lane detection work on curves? let's capture some data when steerign angle is above X and speed is above Y", or "We have a new NN label 'zombie', lets add a trigger that captures first 2 occurrences of this label from a bunch of cars, and then manually sift through those to see if there are any false positives" (cannot even do false negatives with this approach!)
The closest to "new" it gets is the intervention triggers - "the car is on autopilot and XXX and the driver applies brakes/steering/acceleration"

Do we know what all the triggers are? I am wondering how Karpathy got his various snapshots of rare objects, vehicles, lane markings, traffic lights, etc. Purely random triggers?

I am not even sure this is possible. How do you (system) even know there is an object if it was not recognized?

Good question. I know that object detection and object classification are discrete tasks. I also know ConvNets classify individual features of an object before classifying the object as a whole. Maybe a ConvNet could detect a feature of an object, and hence know an object is there, but not be able to classify the object. For example, it might spot a weird vehicle (example) it hasn’t seen before, and detect wheels (a known feature), but have very low confidence classifying it since it doesn’t match up with any previously labelled example. I don’t know. I’m just speculating.

The problem is you don't know what's helpful and what's not and things that are not helpful NOW might become helpful later.

If there are intelligent ways to filter out novel data from non-novel data, that would be a way to do it. The trick is doing that reliably. There may be false negatives for novel data.

Edit: Apparently this is known as “novelty detection”, and it’s an active area of research!

 

[verygreen]

Do we know what all the triggers are? I am wondering how Karpathy got his various snapshots of rare objects, vehicles, lane markings, traffic lights, etc. Purely random triggers?

We have a pretty good idea what the triggers are. I have seen quite a few examples.
I don't know what rare objects he's got, but some might have been misdetected by something. There are triggers for "strange road markings" and such too.
Some might have been captured by their dev vehicles? Non-random triggers might have gotten them while aiming at something else and then highlighted by human labellers too.

Good question. I know that object detection and object classification are discrete tasks. I also know ConvNets classify individual features of an object before classifying the object as a whole. Maybe a ConvNet could detect a feature of an object, and hence know an object is there, but not be able to classify the object. For example, it might spot a weird vehicle (example) it hasn’t seen before, and detect wheels (a known feature), but have very low confidence classifying it since it doesn’t match up with any previously labelled example. I don’t know. I’m just speculating

I have not seen anything of the sort. Triggers allow for a single NN label to be matched in a camera frame (all examples I have seen were matched against the "main" one). I guess you could do several, but I doubt they have low level labels like "wheels".

Internally they might have "vehicle" vs "car" vs "truck" vs "pickup" vs "tank" vs ..... labels, I guess, but that's pure speculation why would they even need to distinguish them in such a fine-grained manner.

 

[strangecosmos]

Do you have access to the full list of triggers, or do you only notice triggers when they have been activated in a hacked car? I am wary of concluding that what hackers have seen is all that exists under the hood.

Karpathy said in his talk that they distinguish different vehicle types, and that it’s important to do so — I would guess for prediction (e.g. a tractor will move more slowly than a car). The examples he used were unusual rare vehicles, though.

“Strange road markings” is an interesting label. How does a ConvNet detect strangeness?

In my understanding, when any ConvNet classifies any object, it does so by classifying the features. So, if a ConvNet classifies an object as a car, it does so by classifying the features of the car — the wheels, the windows, the shape of the roof line, etc. So, a ConvNet could classify a feature without being able to classify an object — hence it would know an object is there, but not know what it is. That’s one way you could detect an unknown object.

Apparently, there is active research in deep learning/computer vision on the problem of “novelty detection”!

 

[verygreen]

Do you have access to the full list of triggers, or do you only notice triggers when they have been activated in a hacked car? I am wary of concluding that what hackers have seen is all that exists under the hood.

I don't have access to full list of triggers, but I've seen quite a few of them at this point to get a good idea.

Karpathy said in his talk that they distinguish different vehicle types, and that it’s important to do so — I would guess for prediction (e.g. a tractor will move more slowly than a car). The examples he used were unusual rare vehicles, though.

Obviously they do at least some vehicle type distinction for obvious reasons. They also track pedestrians vs cycles vs motorcycles because those behave differently too. But "unusual rare" is then confusing and might not mean what I would mean by those words.

“Strange road markings” is an interesting label. How does a ConvNet detect strangeness?

No idea.

The trigger is a label on the main camera, here's an exmple from mid-November 2017:

{"query":{"$if":[{"$and":[{"$gt":["@LegacyDebug.veh_speed_mps",5]},{"$lt":["@LegacyDebug.veh_speed_mps",80]},{"$gt":["@VisionImageEmbeddings.main.timestamp_ns",0]},{"$eq":[{"$mod":[{"$round":{"$div":["@MiscTriggerFields.wall_time_ms",20]}},4]},0]}]},{"$gt":[{"$sum":[{"$labelboost":"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","version":"0.1","camera":"main"},-0.005908]},0.5]},false],"true-stateless":true},"topics":["main","single-frame","latest-frames","vision-rtdv"],"requester":"img-lb-unusual-lane-lines-0072","max-requests":10,"timeout":1510977600,"min-period":30000,"request-time-to-live":432000,"request-wifi-hold-time":432000,"num-requests":0,"previous-request-time":0}

So that implies there's a separate label for that.