I do wonder if perhaps Uber made the video darker before releasing. Or used a camera that is not the main sensing camera. Because I would hope the cars are smart enough to increase ISO to get more detail at night time.
2) They should be using a dark adapted / night vision camera and probably are, the video feed presented is just that, the best one that "presents".
3) They should be using thermal infrared to spot living things
4) They were severely over driving their headlights if what we are seeing is "reality" as seen by the onboard computer.
I want to know how much access to the hardware Uber had after the accident, including physical and remote. I also want to see the streaming logs and full provenance.
(4) is the key point here, if that's the only meaningful telemetry, they're going way too fast. If you can't see then you can't stop. This is still a computer failure.
There is no way, full stop, that Arizona regulators allowed a driverless car with only a single dashcam quality level camera as telemetry to drive around. Either Uber broke the law and let a car on the road that doesn't have working LIDAR and IR and possibly RADAR or the software just doesn't work.
These would have been the 1st battery tests by the car company, surely? Child chasing ball is obviously important scenario. Lidar/sw didnt recognize human/bike combo? But blob was moving & system failed to see or avoid moving blob.That makes me think there should also be a pretrip system check like airlines.
What controls does Uber have that the "safety driver" is doing their job? When the computer gets confused or its computed probability drops below a threshold, when and how does it alert the driver. How many of these self driving testbeds have a person in the seat just for CYA and not actual safety ?
I'd like to see Uber's logs of all the other pedestrian and vehicle near misses where the computer took corrective action.
What testing did Uber do on a closed course with adversarial conditions? Ball rolls into road with a child following, pedestrians walking under a flickering street lamp, people dressed in costumes, a parade, a protest, service workers, a pickup truck with a lost load.
I don't think that would work. Driving is not a fully conscious, deliberate see-analyze-act loop. A lot of your control movements are subconscious, and you expect instant, correct feedback from the vehicle. To see what I mean - if you've ever played a racing game, try opening a racing replay/Let's play video on YouTube, pretend for a while that you're in control (use your WASD keys), and see how your brain complains about your inputs having no impact on reality.
> 3) They should be using thermal infrared to spot living things
To my knowledge the fast frame rate (>9 Hz) thermal sensors are still ITAR-restricted, seriously limiting their application in autonomous vehicles among the other things.
> They were severely over driving their headlights if what we are seeing is "reality" as seen by the onboard computer.
This is the salient point that I was looking for.
Interestingly, US and European headlights standards are very, very different and this accident is a clear example. US headlights are intended to "illuminate the roadway", whereas European headlights are intended to "illuminate obstacles on the road and in the margins". That's not just a wording difference, rather the part numbers for US-spec and European-spec headlight assemblies are different.
Hopefully the NTSB has direct access to all the other telemetry and will release it as well. It would be greatly beneficial to the public if we could see the LIDAR and thermal logs as well.
It's Arizona. I just checked. The high today was 92F. Pavement would almost certainly be warmer than that. How's a thermal camera supposed to detect 98F on a background that is proably +/- 5 degrees of that?
Thermal cameras care about more than just the absolute temperature of an object. The emissive and reflective properties of the material (in infrared) also matter significantly. A human being would almost certainly be visible in a thermal imaging view of that scene.
When you walk outside in the daytime, everything is bathed in a uniform light from the sun, but you have no trouble distinguishing between objects, as they emit and reflect light differently.
Probably not relevant to conditions at 10pm on Monday, when it was between 70F and 57F (the range for 6pm Monday to 12am Tuesday.)
> Pavement would almost certainly be warmer than that.
Sure, when it was 92F the pavement would probably be 140+F.
> How's a thermal camera supposed to detect 98F on a background that is proably +/- 5 degrees of that?
At night, the pavement would be much cooler than a human; at the high temperature you report, it would probably be much hotter; there's a place in between where the problem you have would be occur, sure, but it's neither at the high nor, more to the point, in the conditions when the accident occurred.
See comment above. We routinely walk around (and avoid objects) even though the entire field is uniformly bathed in the light of a single light source.
If you redefine the temperature scale so 90F is 0, then that difference between 98 and 92 becomes the difference between 8 and 2, which suddenly looks a lot more significant. So I think your concern is based more on the arbitrary 0 of the F scale.
If a thermal camera can't detect a person on the road, then a thermal camera is the wrong camera to use. The exact technology isn't important - the important thing is, the vehicle failed to detect a human being crossing in front of it.
Even lows are still in the upper 70s, and again, asphalt will absorb heat during the day and then release it at night. 10pm isn't exactly "middle of the night", either, it's only 3 hours after sunset.
There would still be situations when you'd be attempting to detect a human with a hot asphalt background. Such as a crosswalk at the base of a hill (C is the car and X is the crosswalk):
I wouldn't be one bit surprised. The antics of this company are well known. This incident has my wondering: of all the companies working on this, why is Uber trusted to be testing this technology?
> Levandowski seemed to struggle in other ways as well. In December, Uber dispatched 16 self-driving cars, with safety drivers, in San Francisco without seeking a permit from the California DMV. The test went poorly—on the first day, a self-driving car ran a red light, and the DMV ordered Uber to halt its program in the state.
> The company suffered further embarrassment when a New York Times article, citing leaked documents, suggested that Uber’s explanation for the traffic violation—that it had been caused by human error—wasn’t complete. The car malfunctioned, and the driver failed to stop it.
> The misdirection came as no surprise to the Uber employees who’d spent time at Otto’s San Francisco headquarters. Someone there had distributed stickers—in OSHA orange—with a tongue-in-cheek slogan: “Safety third.”
In 2016 they tried just turning their autonomous vehicles loose in SF without regulatory approval. They got shut down. Rather than apply for permits, they headed off to Arizona, where the then-new governor welcomed them with open arms, an "anything-goes approach", and a lot of anti-regulation blather: https://www.nytimes.com/2017/11/11/technology/arizona-tech-i...
Culture takes time to change and we don’t know how hard they’re trying. A new CEO matters long-term but lower level management matters more in the short-term.
That's not what I was suggesting either. We had a long discussion in a nearby thread about this; I got tired of it but check it out if you're interested.
Trusted in general. Trusted by people. Trusted by authorities.
I mean, there ought to be the point at which someone says, "this company has a long, consistent, documented history of antisocial behaviour and complete disrespect for law, therefore they shall not be allowed to work on this society-changing and life-critical technology". That this doesn't happen is, I feel, a failure of our society/regulatory apparatus.
I’ve gotten 2 products UL certified, and I don’t think it’d be very useful for self-driving cars.
The process seemed to be mainly reactive: only things that had caused frequent problems in the past were part of the standard.
For instance, plug-in devices must have fuses. The standard gives no guidance about the right size of fuses, but a reasonable engineer would of course pick a good size. The standard’s main effect is to avoid cutting safety features to save cost.
We shipped a mobile office robot. We sweated the details of it not running into people or crashing, but the UL cert only verified that we had fuses and flame-retardant plastics and the like.
> I do wonder if perhaps Uber made the video darker before releasing
This video was released by the police department.
Are you suggesting someone from Uber drove down to the scene at 10pm, somehow managed to grab the SD card from the dashcam with all the police there, ran back to their car, uploaded it to their laptop, fired up Premiere, edited the brightness, downloaded the modified video back into the SD, ran back to the scene with the tampered evidence, and only then the police got to it?
Sorry, but if this isn't a conspiratory theory, I don't know what is.
I'm hoping the police and NTSB have guards on the impounded car and they have a strict chain of custody on anything Uber accesses.
To be fair though, Uber can probably extract, and potentially even change this data, remotely. Given their track record, that's not outside the realm of possibility at all. They'd be stupid to attempt it with the world staring at them, but this is also the company that had Greyball, the hell map, blatantly disregarded laws in many countries and had to hire Eric Holder's company to run damage control.
It's probably easier to just enable ssh on the prototype system to enable arbitrary remote debugging than to specifically design a locked down API for accessing certain data.
I was under the impression this was just footage from an off-the-shelf dashcam. I could figure out how to get video out of mine without an instructions manual. Is that not the case here?
You have no idea where the video came from. Completely conjectural example: uber contacts police (who haven't even tried to find any sd card or download the contents or figure out how to convert it to a format postable on twitter) and says "hey I have the video for you."
Did anyone bother to go to that exact spot at night to see what things are like from a human perspective? That strikes me as the first thing to do as a news agency...
The issue is more with the dynamic range of the sensor than the ISO. Increasing the ISO would blow out the highlights long before providing sufficient detail of the shadows (not to mention the increased noise).
Regardless of this camera's performance the LIDAR and other sensors should have picked this up.
Is there any reason it wouldn't be possible to employ multiple cameras each with varying gain/ISO/aperture/exposure/shutter-speed to combat the narrow dynamic range of the individual sensors? Basically create an HDR stream in parallel instead of in series by varying the settings from frame to frame.
Disclaimer: I have a frustratingly poor understanding of this subject, something I desperately need to remedy.
This sort of method is used in the Canon DSLR custom firmware Magic Lantern for HDR video - it brackets the exposure by ISO. However you have some problems with moving subjects: https://youtu.be/5me5jEr4ldQ?t=415 (warning - flashing images)
That's interesting but it still seems to be utilizing a single sensor set to different exposures in series, not multiple synchronized global shutter image sensors each set to different exposures. While such a setup may have little to no use for human perception what I'm talking about is feeding each stream into a vision system, be it a classical vision pipeline or deep net system.
Those cars have multiple sensor feeds, many of which are far better than human vision. I have experience with off-the-shelf IR illuminators, and even an array of half a dozen of cheapo ones improve low light object detection at distances of upto 40-50m by orders of magnitude.
So I feel Uber would have selected whichever feed best supported the explanation that it was "impossible" to avoid such an accident, and handed it to investigators. Given the number of people agreeing with the "impossible to avoid" explanation even in a tech-savvy crowd like HN, I'd say the PR strategy's worked well and saved Uber some more bad press.
I think it's possible to look at the video file and find markers for modification. Certainly this is the case for JPEG, TIFF, PSD, vs Raw files. It's not impossible but very impractical to render a Raw file, edit it, and then unrender it back into a Raw file. It's also pretty malicious and I think it would show a corrupt intent to modify a Raw image in this manner.
There are Raw equivalents for video, but consumer cameras are all post rendered, and heavily compressed. So modification is going to change all kinds of metadata, but in particular there will be more than the expected quantization from double compression.
Again, if that were demonstrated, I think it would show a corrupt intent. The intent to deceive.
Redcode Raw is an example, I'm sure there are others. This doesn't mean there's zero in-camera processing, it just means it's not rendered, ergo it's scene-referred or camera-referred. There can be lens and sensor artifacts removed, like dead pixels.
OpenEXR is another such format that's camera agnostic, not totally dissimilar to Adobe DNG vs the (proprietary) Raw format used by a particular make/model of camera.
Doubtful. There's no way they are making the CCDs/MOSFETS/etc. themselves, you'd need a heck of a chip fab and for almost no increase in usability. Most chips these days are very good at what they do. The main issue is what kind of sensor are they using. However, based on the compression that twitter is using to send you the video, we really don't have any way of knowing what the sensor is (bit-depth, frame-rate, dwell time, wavelength sensitivity, etc).
I don't mean they made their own CCD chips, (and I guess I didn't mean raw like what the voltages are at the digitizer), but the configuration and pre-processing prior to being fed into their algorithm probably is considered valuable enough that they wouldn't want to publicly display it.
You can't increase ISO without washing out the areas in the street lights. One of the difficulties not being discussed here is that this is a terribly lit road, with large dark voids between the street lights. And the biker was crossing in the void.
It is not a terribly lit road. You cannot increase ISO, but you CAN increase dynamic range. The human eye, as an example or as another example a much better camera, can easily see those seemingly large dark voids between street lights.
This is a common theme to all the luddits and uber haters, and I don't get it.
The video doesn't show the rider until less than a second before the collision. Yet it's an article of faith among you guys that this video which does not show an easy path to avoidance actually does, due to various magic incantations:
+ The road lighting wasn't bad. A real eye would have shown something different that the camera can't[1], therefore the video proves what it doesn't show.
+ LIDAR and IR should have shown that, therefore Uber is hiding something because the video doesn't show data that must have been present.
I just don't get it. I'm watching this video and seeing what is clearly a huge tragedy and a near-unavoidable collision.
[1] This is not at all true in the real world, by the way. But I see little value in arguing what is clearly a point of faith and not reason.
Not a luddite here (though definitely an Uber hater). I'm really disappointed by what this accident tells.
IMO this was still a hard case for a human[0], but it shouldn't be a hard case for LIDAR-equipped self-driving car. They were not doing Tesla-style stunts with "we'll drive only with a visual-spectrum camera". They had a suite of sensors on-board.
I expect it'll turn out that either it was Uber-specific problem (e.g. software bug, or lying about LIDAR capabilities, or both), or a deeper problem in sensor tech for self-driving cars in general. The second case would be really disappointing, as - given all that has been published so far - one would expect self-driving car tech to avoid accidents like this.
--
[0] - I mean, really, a pedestrian going through a road like this with zero reflective lighting? Is it even legal in US? In Poland it isn't.
In the wake of the Tesla accident, there was this discussion about the difficulty of detecting stationary obstacles (too many false positives, so they're readily discarded). As here the movement was perpendicular to the movement of the car, I wonder whether that had any impact.
What do you mean by unavoidable? Do you think a human would have made the same mistake in most cases?
Because then my question is, have you ever used a camera at night? It can't see as much as you (at video rates).
What don't you get about people discussing other sensing technologies? Just because it wasn't clear from this video and this camera, there are other sensors which don't require illumination with visible light. Even a front facing radar might do a good job of slowing down.
I'm saying that the evidence as it stands doesn't remotely support the idea that this accident was avoidable for either the driver or the automation, and that I'm shocked at the irrationality with which people are trying to pretend it does.
If you want to make a point about cameras or eyes with evidence, do so. The evidence in the linked video supports there opposite of what you claim.
Here's what the video helps understand. The pedestrian wasn't hiding behind, say, a bush and leaped out. It seems like the pedestrian was unoccluded for the whole time they were on the street.
"Seems like", that is, because the pedestrian is not actually visible! This is the irrationality part. You are literally saying that she must have been visible because you can't see her on camera.
Gah, but she was in the dark. You literally can't see her. You're asserting stuff about human vision being better without evidence[1], and then using an inference from the video that clearly shows this woman was invisible to "prove" your point. Which is insane.
[1] Again, this is just not true. Cameras sensors have HIGHER dynamic range and ISO bandwith than eyes. What is true is that people have better AI behind the scaling decisions and can search better across wide-dynamic-range environments than typical cameras. But again, that presupposes that the driver would have been looking for the black-clothed biker walking a bike without reflectors in the dark shadows between street lights. Which, wait for it, NEEDS EVIDENCE.
By visible I meant visible to active sensors like radar and lidar. This is the second part of my original response to you, and why I then said the pedestrian was unoccluded. Sorry that was unclear. Also sorry you're so frustrated.
Street lighting is a side issue, and the lack of street lighting is AZ is a feature. It's awesome, I get to keep my night vision, and we have the stars:) Making things worse to accommodate confusion about what computers are capable of is a bad idea, and the correct sensors don't care anyway.
It's great that Uber changed the CEO and he's trying to change the culture. But that doesn't happen overnight. This is a 9 year old company with 12000 employees. It has a culture that has been built and reinforced over that time. As CEO, he doesnt make decisions at the front line directly, and it may take him a few years to convince everyone in the company to adopt his rules (and he may even fail, depending on the incentives inside the company). There's no magic wand a CEO can use to change a company overnight.
Are you joking? Uber has always flaunted laws. Even in their home town it was illegal to offer a taxi service without a license. They did it anyway. Later, the city decided to differentiate between ride hailing via app and sticking your arm out to hail a cab. That change happened after uber broke the law.
Third comment I have seen by you mentioning the new CEO, so I think this is worth stating. Does a new CEO automatically change all the thinking from all the employees that were hired under the old CEO? Uber has a compromised integrity because they, as a collection of managers, made immoral/illegal decisions. So yeah, hiding information to save their ass is something I see as possible.
Not just their thinking.. after 9 years, they have policies, practices, rituals, incentives, etc (both written and unwritten) that supported those immoral/illegal decisions.
And the employees aren't helpless either. If they don't like this new CEO, they can choose to withhold information, don't write things down that are no longer approved of, etc while their direct manager looks the other way. They can collectively undermine the CEO until he's completely ineffective and forced to leave.
The culture is one of the hardest things to change inside of a company once it's established. It's self-reinforcing.
So we have to wait a couple of years to see if he's successful. There's no guarantee he'll win just because he's at the top.
I don't know about you, but if I was in his shoes coming in trying to fix a company with a compromised integrity like this, I would be making it very clear to the employees on day 1 that anyone who continued the same old crap or showed similarly questionable judgment (whether to the company's benefit or otherwise) would be fired instantly, end of story. I'd bet that would change things mighty quick. And I don't think this is the only solution either. So yeah, I think it's very doable if the CEO is serious about it, and he very much seems to be. I'm more than open to see evidence to the contrary, but until then, I think it's quite wrong and unfair to assume they would be continuing their previous practices.
Culture is much harder to change than you're making it out to be. The employees are not his subjects that can be bossed around and ordered to behave in a certain manner. That's old school management thinking (from 50+ years ago) that's very ineffective.
The culture is one of the hardest things to change once it has been established. It prevents companies from entering certain market segments; from competing in certain ways; from taking certain actions. In a large company, it's so difficult to change, that it's easier and more effective to establish groups outside of the company to pursue new initiatives.
For example, this is the reason why AWS is separated from the retail segment inside amazon. Why Walmart won't compete with Nordstrom. And why large companies regularly create subsidiaries to innovate on new ideas or pursue new market segments (and they're often established off-site, so the corporations culture doesn't carry over).
GM makes a great example about how difficult that can be. They spent billions on Saturn as an experiment to build a new car company. Later they tried to integrate Saturns success into GM by adopting their culture and other practices... but it didn't matter what the management wanted... Saturn died and GM didn't change. Saturn's culture didn't infect GM--GM successfully killed a culture that tried to change them, even after seeing the success of Saturn.
> The employees are not his subjects that can be bossed around and ordered to behave in a certain manner. That's old school management thinking (from 50+ years ago) that's very ineffective.
Why the false dichotomy? How is telling your employees "you will be fired for lack of integrity" equal to "bossing them around"? This is quite the straw-man and it undermines the rest of your points for me. It's not like the only 2 options out there are to order people around like your servants or let them do whatever illegal hell they want. He could be strict about the red lines on integrity without ordering them around like his servants with regards to general management. Why do you make it seem like the only possibilities are the extremes?
What is "lack of integrity"? It can't be defined granularly enough by a CEO. He can make an example of some particularly egregious behavior--and that does help--but there are millions of small decisions that happen every day.. and the employees will need to decide if each of those actions represents a "lack of integrity" or not. And the truth is, they already have some practices, that they aren't going to question unless directly prompted to do so. Just as you repeat numerous rituals and habits regularly--you don't reexamine them until something forces you to.
What this means is that the CEO cannot order everyone to behave with integrity. They must be convinced of what integrity is, what type of behavior that represents, etc. So that when they make these small decisions, they are consistent with the culture the CEO is trying to create.
You can't order them to behave with integrity--because that's so vague as to be meaningless. Everyone will make up their own definition of integrity, and it'll always fit so they turn out to be right.
So I don't see the false dichotomy... There's no order a CEO can give that will change the culture in the way you describe (at least not in a large company like Uber).
(I think it's worth pointing out that the people inside of Uber who made the bad decisions likely didn't see them as bad or wrong. Their mission and culture supported those decisions, and they believe them to be right. So when you say 'behave with integrity'... that's meaningless, because they believed they were right when they made the wrong decision initially. They were already acting with what they believed to be integrity.)
Yeah, it's impossible to define it. But thankfully, you don't have to unambiguously define ethics and integrity for all of humanity and posterity to get somewhere useful. We're not writing probably correct algorithms here, we're dealing with humans. Some linear combination of "something illegal", "something you don't want landing on the New York times next to your name", "I know it when I see it", and "if this still isn't clear for you, you test me at your own risk" would be sufficient to take care of it: either they'll figure it out by themselves or they'll find out the hard way.
The clarity of that rule only exists in your head. Each person will draw the line differently, assess the risk differently, etc. Saying "I'll know it when I see it" does not result in a culture change. That just results in the CEO randomly enforcing his rule when he finds out about something he doesnt like.
Again, like I said, it doesn't need more clarity than however clear or vague it is at the moment. Either people will figure out how to stay away from the gray area and whatever they understand your red line to be, or they won't and will instead have to find out the hard way. Your mere assertion that this won't work isn't any more convincing than mine that it will work.
Well actually in management it is well known that your method wont work. There's a history of your approach, a lot of research and things we've learned from corrupt corporations, etc and there's a reason it's not used anymore. History, methods, purpose etc of corporate culture is quite interesting, and I fully encourage you to read more about if it interests you. Uber's new CEO has a monumental task ahead of him.
You're arguing they deserve the benefit of the doubt until proven otherwise. Most people seem to feel they've done enough damage that it's on Uber to prove they're worthy of trust.
The cars don't have to increase ISO to get detail, that's for human eyes. Increasing ISO only works because we can't see details in shadows, but computers don't have that limitation. A neural network wouldn't have to add 100 to each pixel to make out detail, like a human would.
> The cars don't have to increase ISO to get detail, that's for human eyes. Increasing ISO only works because we can't see details in shadows, but computers don't have that limitation. A neural network wouldn't have to add 100 to each pixel to make out detail, like a human would.
I'm not an expert on this stuff but I think you're misunderstanding or mixing up the ISO value with exposure compensation. The ISO value corresponds to the sensor's sensitivity, and even RAW photos (which are more or less dumps of the raw sensor data) are captured with specific ISO values, and look different at different ISOs.
i) what human eyes need to do to get details from a dark image (increase ISO and risk washing out bright areas); and
ii) what human eyes do when looking out at the real world (rely on the huge dynamic range that the human visual system is capable of, which dynamic range is about an order of magnitude larger than any video camera sensor);
It's worth pointing out however that humans can't make use of that whole order of magnitude range at the same time. The visual system has to gain up and down as the average light in the field of view changes. This takes several minutes or even a half hour to fully dark adapt.
Sorry, in short you're wrong, yes they do. I'm on my phone so I will be brief.
The ISO determines the sensitivity of the sensor in the camera to the photons hitting each bucket in the "sensor array". This sensitivity combined with frequency of the photon hits determines the data that makes up the actual image. If you could theoretically dump the RAW data from an image recorded for a different ISO for the same image, it would be different for each one.
You can't "boost" an image's ISO post-capture, because the sensor data is already captured. What you can do is increase the values of those pixels that are appreciably close to buy not equal to zero, but for which data exists, giving the impression of pulling out data from the shadows, but the limitations/complexities are too much to go into here. When you do so in practice, you do not have as much relative detail in those extremes.
What we interpret as "noise", typically in high ISO low light photography, is a real thing that the computer cannot perfectly compensate for, because it's a fundamental property of both the hardware, but more importantly, of background light/radiation and the probability/quantum nature of light...
