There is definitely a negative aura associated with this term in traditional development, so let's start by saying this is a new and improved iteration of estimation accuracy for agile development. In this version, nirvana is not 100% (i.e., actuals = estimates), but instead to measure consistency over time.
Similar to velocity, estimation accuracy is a way to measure a team's pattern of work over time. At the beginning of each iteration, teams try to identify and estimate the work for the iteration. Rarely, if ever, will they define all of the tasks for an iteration. In addition, they will estimate these tasks. [As a special note -- An estimate is exactly what its name implies, a projection based on the information available at the time. If you asked the team to identify and estimate the tasks for the iteration half way through the iteration, I bet they would improve (so we should do this, right...?).]
Anyway, through the life of an iteration, the team will spend time completing the tasks, adding new tasks, collaborating, etc. At the end of the iteration, they will have worked a certain amount of time. Over time, the ratio of estimated time vs. actual time will often stabilize. For example, on some teams the amount of time worked may exceed the original estimate time by 20%. In another example, a teams actual time may be 2x the estimate. Neither ratio is bad, particularly if these ratios stabilize over time. Estimation accuracy, however inaccurate, if consistent, can provide tremendously valuable insight. If the 2x teams typically operates under similar ratios (say 1.7x to 2.3x), then when they estimate an iteration, I know we will typically expend twice the estimate. If they estimate a project, then I may even predict with some level of confidence that it may cost twice an estimate (plus some percentage contingency).
One path might be to try to improve estimates over time, but another path would simply be to value this consistency and leverage it. Although someone might be incorrect, if they are always incorrect by 100%, then I can always be right.