Comments by "Luredreier" (@Luredreier) on "Why This Self-Driving Tesla Car Hit That Truck | Bumper 2 Bumper | Donut Media" video.

  1. 11:51 Neural networks tends to work best when you have few inputs and outputs to deal with. I don't think adding a lidar would have helped much, or if it's too good the neutral network might have ended up relying on it too much causing it to miss other things.
    1
  2.  @zihechen3111  LiDAR is not stupid, there's things it shows that a camera won't pick up. A camera essentially is dealing with a 2D grid of information that has no inherent depth information contained in it. With multiple cameras you can kind of guess a distance, but cameras won't always be enough. Especially in poor visibility conditions. Radar LiDAR etc contains information about depth and works better in poor lighting conditions. Engineering is complicated, and usually when different engineers come to different solutions you're dealing with situations with different tradeoffs and many complicated considerations. So calling either option "stupid" just isn't helpfull at all...
    1
  3.  @zihechen3111  Try using headlights in a Norwegian winter storm, I dare you... As for ultraviolet and infrared, that's still colours close to the visible spectrum and had the same flaws in that they're absorbed, dispersed etc by the same or similar materials to visible light. A LiDAR is essentially just such a "headlight" except far stronger and focused on a point instead of all over the place so you increase the likelihood that some of it gets through and since there's so much more of it there then anything else it's easier to pick out from the ambient lighting so you can do essentially the same trick as in ecolocation and measure the time between a signal being sent and received, something regular headlights are too primitive for and also something they're not useful for even if they had the tech because that kind of light would be difficult to distinguish from ambient lighting... And if you start adding night vision etc you may as well start looking at radar and LiDAR anyway as you're adding the same kind of extra input for the neural network to compute... It's a tradeoff. If you add more inputs you'll need a bigger network and more training and your overall outcome might not be better. Yet it's more *resilient*.
    1