Hacker News Reader

Comment by daveguy | Hacker News Reader

A direct reply from Elon Musk on twitter about why radar did not recognize the white side of a trailer across the road when the camera missed it:

"Radar tunes out what looks like an overhead road sign to avoid false braking events"

https://twitter.com/elonmusk/status/748620607105839104

EDIT:

When the "overhead sign" comes down below overhead clearance of the vehicle the signal should not be masked. There should have been some braking action in this case. If there was not then the tesla autopilot is unsafe. This is the same blind spot discussed a few months ago that caused a tesla to run into a parked trailer using summon mode:

https://news.ycombinator.com/item?id=11677760

This seems like a serious flaw in autopilot functionality. Trailers are not that rare.

I would be interested if "autobrake/autofollow" functions of other car companies have similar problems.

View on HN

Also an interesting comment:

This is why trailers require safety guards in the EU. Cars simply cannot get underneath them.

https://twitter.com/jn_cn/status/748626793180069888

A reply to that tweet:

"our new Prius applied the brakes while driving under an overpass. I'm glad no one was following too close"

The Prius uses cameras not radar, and I've driven under hundreds of overpasses without incident.

More likely someone confusing the brake warning with the auto-braking shutdown warning. Because it uses optical cameras, driving into bright sunlight or into deep shadows can cause the system to shut down momentarily, and it beeps to let you know.

Even best case scenario that tweet has nothing to do with this incident because the technology and systems are different. Worst case it never applied the brakes at all, and the driver got confused.

The radar might not be able to clearly distinguish between objects at the height of the cab and objects at signage heights. Just about everything on the road is mostly on the road.

If the system can bias its sampling towards the ground, it would make a lot of sense to do so. Lump "above the ground" and "very above the ground" into the same category and use other sensors (for example, the camera..) to detect the edge cases where it matters.

I agree it probably can't distinguish between those heights, which is a danger either way : not recognizing going under a trailer or accidentally braking at overpasses. Neither are very safe or pass "human level".

Of course we should have undercarriage bars regardless.

It also seems strange that the rest of the tractor trailer was not picked up by the radar when it was entering the road in the first place - the front part that is pulling the trailer is typically lower and should not have been mistaken for a road sign or otherwise.

It may have. AI isn't quite to the level of understanding "object permanence" yet. So, if you would let a 2 year old drive you should be fine with autopilot. The more I think about this the more I think it will set self-driving back 10 years.

Or that radar cannot detect an object moving perpendicularly in front of you regardless of overhead clearance. I assumed that auto pilot was more advanced than what is now being described. They have a lot of work to do here.

Seems odd to me that the autopilot doesn't use them together, as it would seem that the autopilot would easily be able to see with the camera whether it was a sign, as at least in the United States they seem a somewhat uniform green.

I think the likely answer is Tesla "autopilot" is a glorified cruise control + lane assist + data collection for future intelligent controls.

Google probably has close to this capability, but they are at the leading edge of image recognition.

In order to depend on it the AI would have to be able to distinguish 99.9+% and would have to be able to tell the difference between a green truck and a road sign. I honestly think self driving cars will require intelligence indistinguishable from artificial general intelligence.

This was state of the art in 2014 (improved in 2015, but this paper is on arxiv and gives the idea):

http://arxiv.org/pdf/1409.0575

Even top scoring Google had classification errors over 6% of the time in 2014. Object localization was much worse. Even if it is at 99% that means 1 of every 100 object will be misidentified. You probably identify that many random (non-car, non-sign, etc) objects in a week driving, maybe in a day. That kind of error is unacceptable in self-driving cars and it is state of the art. That is one of the reasons they are still testing and not selling self-driving cars at Google.

Better than human may not require quite 99.9+% object classification, but I do wonder how great the self-driving car records would be if someone wasn't always there to take the controls. It certainly wouldn't be "better than human" at this point.

Yes, Tesla system is a fancy lane following dressed in powerful PR.

Btw I dont feel Imagenet results are relevant, it only tests static 2d images. Car system can have multiple cameras providing 3d reconstruction in near realtime. Mobileye (what tesla repackaged) probably has only one front facing camera and is too stupid beyond 'hey this uniformly colored blob is in my lane'

I was putting image net performance out there as a bare minimum capability. I agree 3d real-time reconstruction is harder (although aided by multiple cameras + radar, etc). You're right, it probably isn't a fair comparison because of the convolution of real-time needs and additional sensors/cameras available.

I meant the opposite, its immensely easier guessing shapes on a moving platform with multiple cameras providing 3d reconstruction than guessing exact name of an object from one random photo.

Car system doesnt need to know the name of a dog breed, all it needs is a shape on a collision course. You can do stereo 3d reconstruction on a 10 year old laptop in real time. You can even do 3d reconstruction with one camera in real time as long as objects move slightly perpendicular to camera http://www.robots.ox.ac.uk/~gk/PTAM/ http://vision.in.tum.de/research/vslam/lsdslam.

You definitely need to know more than just "3d blob at position x" because your expectations and avoidance will be different for car, truck, motorcycle, road sign, human and especially for things like "grocery sack" vs "cinder block". Less relevant on the hwy, but it definitely isn't simply an "easier" visual task. It is convoluted with things that make it easier (multiple cameras) and things that make it more difficult (real time adaptation).

Edit: even not needing to know details like "dog breed" it still needs at least a solid 200 object classes to cover not-so-edge cases (which is what had the 6% error rate in 2014).

I dont think so. Walk before you run. There is no difference between a human, bale of hay or a deer when it comes to objects entering your trajectory. Braking in front of all 3d elements obstructing the way is preferable over Tesla decapitation(TM).

Real time is not difficult, we got there 10 years ago for primitive 3d reconstruction at 30Hz on ordinary laptop, today its doable on a phone. It would definitely be better than single camera + pattern matching (current mobileye).

Well, the fact that no one has a product despite billions of dollars in research and at least 5 years directed serious effort (in the latest generation) across every major tech and auto company would indicate that it is a more difficult problem than you think.

There are plenty of commercial products, just look up ICCV ECCV (there is even a dedicated track for autonomous driving http://www.cs.cmu.edu/~CVRSUAD/). But nobody is crazy enough to accept all liability of selling packaged 'autonomous driving system', even MobilEye doesnt do that. Tesla made a big mistake branding their young flawed tech as something it is not.

edit: This is from last years conference, running on a laptop in real time: https://www.youtube.com/watch?v=oJt3Ln8H03s

Reply on news.ycombinator.com