People are bad at estimating, particularly if they're doing it alone, don't have much information or are trying to estimate absolutely. But we're pretty good at estimating relatively, doing it in groups (witness bean counting contests) and even with just a little information we can make improved estimates.
So the ideal task planning in my mind goes like this: the team gathers around the backlog to see what we think the next priorities are. Briefly we chat about each one, enough that everyone has a basic sense of what it is and no one is blindsided. We intentionally don't want to talk about it enough to implement it, just enough to get a sense for it.
The team then collectively estimates the size of each item with a hidden bid system: each person picks a number on an arbitrary scale (let's call it points) that only has meaning relative to the point values of the other things on the list. Once we have a handful, it becomes pretty easy to say "foo is harder than bar, but not as bad as qux".
