Exactly.
People can't get out of the linear thinking.
Well if I have 2 of something than it is better than 1 of something.
That is rarely a given.
But when you add multiple + disparate sensors and trying to unify the data, you have to make some decisions. (just some random thoughts)
- You cannot get them to agree 100% - which one do you take as "right" and in which situations? How do you write logic to make that decision for ALL situations?
- eg LIDAR cannot read traffic light color but knows it is there so you have to rely on camera vision to give you the answer.
- other examples are not that simple.
- You can chose to solve vision with lidar and vision with camera separately
- that is a LOT of time and money.
- do you unify them or run them in parallel? and where / how do you unify them?
- there is no guarantee that the two separate stacks can easily tied together, might require a medium (a NN stack that consumes the output of two vision stacks and puts out one final result)
- now you have 2 vision stacks to maintain, train, fix, etc
- You can do "sensor fusion" and then feed the data into a your main NN stack
- this is what Tesla seems to be walking away from right now...
People tend to only look at the positive of each component, but each and every component has its drawback.
Complexity itself is a big tradeoff!