In this paper, we investigate the decoding of flashed, full-field visual stimuli while recording from a population of retinal ganglion cells. We present a direct statistical method for determining the likelihood that a response was evoked by a particular stimulus, and use this method to estimate stimuli based on microelectrode array recordings in the turtle retina. This method uses the well-known time-varying Poisson model of neural firing, along with extensions to accommodate neural refractory periods. Unlike other approaches commonly used for Poisson processes, the specific formulation presented here is bin free and requires few user-specified parameters. Statistical dependency issues and the effects of stationarity on the estimation method are also discussed.