So, how are they doing this when the public results for recognising pedestrians from images seem to be so rubbish? (Or at least, don't seem to have a sufficient number of nines precision for this purpose). Is it easier to recognise a pedestrian from LiDAR, or is everything just very precautionary?
I see this sentiment but I think there's a flawed, tacit assumption underlying it. Image classifiers might not be very good at calling out a presentation in a static photograph, but an AV doesn't need to classify objects to avoid striking them. For AV purposes, uncertainty about part of a scene is also a useful input.
Recognition is important for predicting future state. Detection is important (critical) avoiding hitting anything. Short of edge cases like a plastic bag blowing in the wind, self-driving cars really don't want to come in contact with anything they detect, even if they can't recognize what it is.
