Two dimensional magnetic recording (TDMR) is a new paradigm in data storage which envisions densities up to 10 Tb/in2 as a result of drastically reducing bit to grain ratio. In order to reach this goal aggressive write (shingled writing) and read process are used in TDMR. Kavcic et al. proposed a simple magnetic grain model called the granular tiling model which captures the essence of read/write process in TDMR. Capacity bounds for this model indicate that 0.6 user bit per grain densities are possible, however, previous attempt to reach capacities are not close to the channel capacity. In this paper, we provide a truly two-dimensional detection scheme for the granular tiling model based on generalized belief propagation (GBP). Factor graph interpretation of the detection problem is provided and formulated in this paper. Then, GBP is employed to compute marginal a posteriori probabilities for the constructed factor graph. Simulation results show huge improvements in detection. A lower bound on the symmetric information rate (SIR) is also derived for this model based on GBP detector.