An Efficient Run-Time Management Methodology for Stereo Matching Application

This paper presents a methodology for Run-Time Management (RTM) of algorithmic parameters. The RTM is able to trade-off the algorithm output quality and the execution time. Thus, once a requirement in terms of maximum execution time is set, the RTM dynamically tunes the parameters in order to maximize the output quality while respecting the given requirement. The run-time decision making relies on design-time modeling techniques able to characterize key relations between algorithm parameters, execution time and output quality. Models generated during the design-time analysis are accurate enough to drive the RTM in its decision making while enough generic to model application behaviors over datasets which were not included at design-time. In this paper the methodology is applied on the Stereo Matching application, a computational intensive artificial vision application aimed at inferring object depths using two or more cameras. Experimental results prove the effectiveness of the methodology which is able to identify high quality solutions respecting required deadline while introducing negligible overhead.