FSD Visualizations!!!

[cypho]

Yes we do, humans can be that good! I think I've only run one red light in my life accidentally (I was really tired and it was very late, and I did notice just before entering the intersection - and I think if there had been cars around I would have snapped out of it earlier). To be clear, I've never run one deliberately, unless it was defective (a next-level challenge for FSD I guess). I think I've probably had something like 150k tests at this point (wild guess). Pretty sure that computers will have to be quite a bit better than me though, I would think at least 100x better.

How can you possibly know how many times you have not stopped at a light that you didn't notice? If you don't notice the light you would be totally unaware that you ran it.

 

[AlanSubie4Life]

How can you possibly know how many times you have not stopped at a light that you didn't notice? If you don't notice the light you would be totally unaware that you ran it.

Ha. That’s a very existential question. How can I be sure of anything? I guess the evidence suggests I don’t habitually obliviously run red lights. But no, I can’t be sure. I noticed when it did actually happen.

It is kind of a life or death thing. So I feel like I would notice. That is all I can say.

Pretty sure I am better than 99.999% though. I’d like to add another 9 but the data does not support it.

 

[powertoold]

I don't understand what you guys are fussin about.

I was wondering if there's any computer vision precedent for 99.99% accuracy at perception in real world objects. I follow ML and CV developments and have yet to see anything with such high accuracy in a wide range of environments.

 

[AlanSubie4Life]

. I follow ML and CV developments and have yet to see anything with such high accuracy in a wide range of environments.

I don’t follow this stuff, but you may be right! If so, it seems like a problem that needs solving still!

 

[Daniel in SD]

I don't understand what you guys are fussin about.

I was wondering if there's any computer vision precedent for 99.99% accuracy at perception in real world objects. I follow ML and CV developments and have yet to see anything with such high accuracy in a wide range of environments.

I sure hope that Waymo hasn't deployed their Level 4 vehicles with only 99.99% accuracy at detecting red lights. If you encounter a red light every 10 miles and don't detect it 1 out of every 10,000 times it's going to be tough to get to the human accident rate of 1 per 500,000 miles and a fatality rate of 1 per 100 million miles.

 

[powertoold]

Waymo has HD maps down to centimeter and all lights and signs mapped out. Sure, they need to read the lights, but there's much more data control and less variability for them. Apples to oranges comparison here

 

[Daniel in SD]

Waymo has HD maps down to centimeter and all lights and signs mapped out. Sure, they need to read the lights, but there's much more data control and less variability for them. Apples to oranges comparison here

It's almost like there's a huge benefit to using maps.
I wouldn't be surprised if 99.99% accuracy for unmapped lights is acceptable since new lights only get added to a mapped area very rarely. If you know about the existence of a light you can assume it's red if you can't detect its state with high confidence and obviously false positives are much safer than false negatives!
Tesla is definitely taking the hard road in trying to do it without maps.

 

[emmz0r]

Waymo also is restricted to a place where it almost never rains

 

[Bladerskb]

Green letting us know we already know.

oh, btw forgot to add, performance in determining hte light color is about matches what was observed on hw2 in march 2019 (i.e. it's not super robust yet as many have noted before me I am sure)​

green on Twitter