Raven is a multi-object adaptive optics (MOAO) demonstrator that will be mounted on the NIR Nasmyth platform of the Subaru telescope in May, 2014. Raven can use three open-loop NGS WFSs and an on-axis LGS WFS to control DMs in two separate science pick-off arms. Centroiding in open loop AO systems like Raven is more difficult than in closed loop AO systems because the Shack-Hartmann spots will not be driven to the same spot on a detector. Rather the spots can fall on any combination of pixels because the WFSs need to have sufficient dynamic range to measure the full turbulence. In this paper, we compare correlation and thresholded center of gravity (tCOG) centroiding methods in simulation, with Raven using its calibration unit, and on-sky. Each method has its own advantages. Correlation centroiding is superior to tCOG centroiding for faint NGSs and for extended sources (Raven open loop WFSs do not contain ADCs so spots will become elongated). We expect that correlation centroiding will push the limiting magnitude of Raven NGSs fainter by roughly one magnitude. Correlation centroiding is computationally more intensive, however, and actually will limit Raven’s sampling rate for shorter integrations. Therefore, for bright stars with sufficiently high signal-to-noise, Raven can be run significantly faster and with superior performance using the tCOG method. Here we quantify both the performance and timing differences of these two centroiding methods in simulation, in the lab and on sky using Raven.