I am a little confused about RCNNs. The first intuition that I got was that we use RCNNs to propose the regions that are more likely to have an objects and then run Object Detection.
I initially though that this implementation would be helpful in terms of computing cost since we are not running our algorithms on a regions that is not likely to have an object in it.
But then, we say that RCNN tend to be slow.
What am I misunderstanding there? What is the advantages of RCNNs?
The drawback of RCNN is “defining regions”, which needs to be accurate to point areas that we need to tackle. It is sometimes time-consuming (rather than convolutions !), and sometimes miss-classify. I suppose there are some improvements, but essentials are those points. So, personally, I did not dig in detail…
