Hacker News Reader

Comment by bnt | Hacker News Reader

Just yesterday I was almost killed while driving back home, because some jackass left a camo-painted trailer in a poorly-lit alley right next to my home. There is no “autopilot” that could spot it in the dusk / sunset. Until Tesla adopts a Lidar, I wouldn’t use the “full self-driving system”

View on HN

Not trying to troll, but your two 'cameras' were able to identify it without LIDAR, why shouldn't two or more of Tesla's / name your manufacturer?

The dynamic range of your eyes vastly exceeds any "normal" camera on the market today and have much better low light sensitivity. Not only that, they're paired to a biological supercomputer whose "programming" is more powerful, efficient, and robust than anything currently in silicon. It's certainly possible to drive autonomously with cameras-only, but it's very much an open question whether we can do so safely today.

And yet with our two cameras, we still get into a lot of accidents. Why would we not want to improve upon that with additional sensors like LIDAR?

So you're going to hold off riding in vehicles with human drivers until they get augmented with LIDAR as well?

To be clear, I am not saying don't use LIDAR. I was calling out OP's statement that they would feel unsafe in a car without LIDAR since it was unable to detect an unsafe condition that they detected themselves without LIDAR.

because Tesla uses 3 front facing Black/Red ~1000x600 36 frames per second cameras mounted in the center with no ability for parallax.

> There is no “autopilot” that could spot it in the dusk / sunset.

Are you saying you were driving on autopilot in a Tesla and it didn't spot it or you were driving a normal car and couldn't spot it?

If the later, you're assertion that because it was hard for you to see a computer wouldn't be able to is pretty strange. There's plenty of sensors that can spot a stationary object in the way even though it's camouflaged to humans, IR being the most obvious.

Not to mention radar, which every Tesla has.

Radar isn’t great at detecting stopped objects especially when the sensor is moving at highway speeds, which is why most AV companies are using LIDAR

The situation described by the GP doesn’t sound like highway speeds. Regardless, most AV companies use a combination of LIDAR, radar, and vision (which you probably already know - just clarifying.)

I'm curious about the details of how a trailer in an alley could almost kill you, were you going highway speeds down an alley?

How close did you come to colliding with it before your eyes and brain prevented this?

I am also confused by this. Alleys (at least in California and Texas) have speed limits of 15 mph.

You don't have a lidar sensor in your head. You have binocular vision. There is no reason we can't achieve parity with human ability with a camera suite. In my opinion lidar is a fools errand. Also I would not expect a tesla to be speeding down an alley.

Perhaps that will be possible someday, but we still don't have affordable cameras that can match the dynamic range of the human eye. Cameras have superior resolution but they struggle when the scene is unevenly lit.

If we’re just designing a system that is just as good as humans, your argument makes total sense. But it looks like we want a system that makes fewer errors and doesn’t doze off and has no alcohol issues. So it’s perfectly acceptable to want something that can also see a little bit better in the dark.

The implication here is that visual object detection in the weak link in human driving safety: I would counter that distractibility and poor response times are the weak link.

The human eye can detect two photons so the brain can identify a star in the night sky. Show me a camera suite that can do that.

I don’t think we can “see” two photons. The brain will filter that away as noise. Cameras definitely can though :)

> The human eye is capable of detecting a single photon, the smallest unit of light, according to a new study.

Google is not your friend, but it searches this just fine.

Did you look at that study? N=3 getting it wrong most of the time...

Reply on news.ycombinator.com