Our representation makes
object boundaries explicit, allowing their structure to be exploited by the
algorithm. Note that it is very
important |
|
to make the structure
explicit. Although inference
algorithms can try to find decompositions on their own, this is a hard
problem by itself. The |
|
algorithm can gain tremendously by having
the structure made explicit by the representation. |
|
This turns out to help a
lot, but not quite enough. |
|