You can install our site as a web app on your iOS device by utilizing the Add to Home Screen feature in Safari. Please see this thread for more details on this.
Note: This feature may not be available in some browsers.
just hope there's enough processing power left for some more AP2 goodies
Due to neural nets being mostly black-boxes, nobody not really understanding why it works, I believe it's best to separate them to keep them manageable. You'd not want lane keep regressions when trying to improve rain detection accuracy. I believe verygreen also confirmed a separate net for this.I wouldn't worry. Neural networks aren't quite the same as other forms of programming, and not all new algorithms have any additional cost. In this case, it looks like the rain sensing is just one(of many) outputs of the same neural network that does anything. I'm going to guess that it was probably needed in that network anyway, even if not used for this purpose, in order to allow the network to learn around the distortions caused by rain.
Due to neural nets being mostly black-boxes, nobody not really understanding why it works, I believe it's best to separate them to keep them manageable
Yeah, those are good points. As long as the problems are enough related. I guess rain sensing can be if it's at a general "Something is obstructing my view", but not sure if the existing neural network is mature enough to know what to expect to see. So my guess is that it just recognizes droplets at this point, and will be improved later.The "not understanding how they work" bit is only partly true. We don't understand the details of *how* a given non-trivial neural network is solving its problem. We do understand fairly well how(and why) the learning process works, what kind of things can help a network learn, what architectures work well for given tasks, etc.
Separating networks is not always a good idea. They should usually be kept separate if their tasks lack any overlap. However, it's fairly common to have a network produce multiple different outputs, because learning about how solve one problem helps the network solve a different, but related problem. I work in the field and have seen multiple different networks produce the *same output feature* because learning to produce that particular feature made the network have better accuracy in its main task.
I see a path here for that to be true for the rain sensing network. The object detection outputs of the network need to be robust to distortions produced by raindrops(or broader areas of water) so the car doesn't suddenly hallucinate an obstacle in front of it. Having it explicitly learn about detecting areas of distortion will likely help keep it robust to that distortion.
For @verygreen, my understanding of what he said is that there is a separate *output* for rain level, not a separate network.
For @verygreen, my understanding of what he said is that there is a separate *output* for rain level, not a separate network.
No, according to @verygreen there is a separate neural network for this - the only NN using the fisheye now anyway... The NN in question is called: fisheye_autowiper
Links here if anyone wants to research: Rumor summary: Blind-spot cameras, Rain sensing, Level 3, Big battery, Interior/HUD
need to be a touch more sensitive to fine drizzle, other than that pretty good.
I wonder if one of hte NN inputs is if the driver prods the single wipe button, as this could be used increase the probability of wipe requirement, thus tuning the system to a driver's preference.
Often thought that AP is a bit "thick" in this repect, as when a driver in the same GPS location daily has to disconnect AP in precisely the same place to avoid a tree mating ritual due to some dodgy road markings, always though AP could learn this, Maybe it thinks it is a superior intellect already.
It finally rained in LA so now we get to join the discussion. I would say it might be overly sensitive. I am leaving mine on the first position and it wipes too often for me. I noticed a pattern. Wipe twice, pause. Wipe twice, pause. It happens in quick succession. Wipe wipe pause, wipe wipe pause. Clearly that’s a condition of the amount of rain but a wipe here and there is plenty in this rain. And if that’s not articulate or specific enough well...need to be a touch more sensitive to fine drizzle, other than that pretty good.
I wonder if one of hte NN inputs is if the driver prods the single wipe button, as this could be used increase the probability of wipe requirement, thus tuning the system to a driver's preference.
Often thought that AP is a bit "thick" in this repect, as when a driver in the same GPS location daily has to disconnect AP in precisely the same place to avoid a tree mating ritual due to some dodgy road markings, always though AP could learn this, Maybe it thinks it is a superior intellect already.
More rain here in Palm Desert!
Finally got a longer more real world test on camera.
They perform perfectly in position 2!!!!
Everything I hoped for.
2+ minute clip Thanks Elon!
Yes, very excited even though it hasn't rained yet and my car clears impending fog mist.... pretty darn impressive.You know when you're a Tesla enthusiast when.... you find yourself watching YouTube videos of windshield wiper activity.
In position 2, did you not expect more frequent wiping given the accumulation of raindrops? Iirc the release notes, position 1 was "medium" and position 2 was more... active, yes?
In any case, nice video.
In my limited testing today, the wipers are certainly operating differently than before when in positions 1 and 2. Can't yet say whether better than the old-style timed/consistent wiping, but it's not worse.