Among the many challenges presented by the future ATLAS detector at the LHC are the high data taking rate and volume and the derivation of a rapid trigger decision with limited resources. To address this challenge within the ATLAS second level trigger system, a Region-of-Interest mechanism has been adopted which dramatically reduces the relevant fiducial volume necessary to be readout and processed to small regions guided by the hardware-based first level trigger. Software has been developed to allow fast translation between arbitrary geometric regions and identifiers of small collections of the event data. This facilitates on-demand data retrieval and collection building. The system is optimized to minimize the amount of data transferred and unnecessary building of complex objects. Details of the design and implementation are presented along with preliminary performance results.