Why delete the history on a classification change?
Shouldn't classifications be tiered? In this case, while the system was struggling to PERFECTLY classify the object, it was clearly thinking it was something that should be avoided (oscillating between car, bike, other).
In this case, I would expect the system to keep it's motion history. IMO, this could have prevented the accident because although it didn't determine it was a bicycle/person until "too late" ... it had determined with plenty of time that it was maybe a car, maybe a bike.
A) Should I avoid driving this car into this thing?
and then there are subsidiary questions that help to answer that, but would be fully obviated if you already had a good answer for A):
B1) Is this a human?
B2) Is this a vehicle?
B3) Is this unrecognizable?
The "natural category" here is "drive into" vs "don't drive into", not "human vs vehicle vs fruit stand vs other".
[1] https://www.lesswrong.com/posts/4FcxgdvdQP45D6Skg/disguised-...
In the article, the A)-type question is whether the object has vanadium, and the B)-type questions are whether the object is a blegg (blue round thing) or a rube (red square thing). The distinction becomes stark when you know it has vanadium, but it doesn't neatly fall on the blegg/rube continuum.
That is, it doesn't really matter if the object history is "deleted" or not; if it can't associate a new data point with a previous history (by identification or predicted position), the practical result is the same as if there is no object history.
This could be a result of using velocity based tracking, which I don't know that Uber uses, but is a fairly standard method, as it's what raw GPS measurements are based on.
This sounds a bit strange, given that the entire technology around which autonomous driving is built is about identifying and recognising patterns in extremely noisy and variable data.
If so, how come these categories where not encompassed into a more abstract "generic object" whose relative position seems to be getting closer since first detection ? That ought to have triggered the same braking scheme as, say, an unmoving car detection ahead.
I'd go for : utter engineering malpractice.
But regardless, taking sensor classification (inherently error-prone) at face value is engineering malpractice.
This might be a little too nitpicky but it doesn't get wiped. It's simply no longer associated with that object because it's considered a different object. It's still a huge, glaring issue, obviously, but all the data is still there.
In this particular case, the "object" was identified as one type of object and so all of the data related to that was classified as "car" info, for example, and then, when it's reclassified, that additional data starts recording to the "bike" info bucket. The software should have been keeping track of certain data regardless of that classification but it's only seeing the latest bucket of data. If the tracking history got wiped, we wouldn't have the data to look back on to see how this was all happening.
Those still all seem to fall into the category "thing you should avoid hitting", though, right?