Estimating how long any given task will take seems to be one of the hardest parts about software development. At my current shop we estimate tasks in hours at the start of an iteration, but once the task is complete we do not use it to aide us in future estimations.
How do you use the information you gather from past estimations to refine future ones?
By far one of the most interesting approaches I've ever seen for scheduling realistically is Evidence Based Scheduling which is part of the FogCreek FogBugz 6.0 release. See Joel's blog post linked above for a synopsis and some examples.