We present an algorithm for generating panoramic images of complex scenes from a multi-sensor camera. We further present a programmable graphics hardware implementation to process the large data sets more quickly. Because the sensors do not share the same center of projection, nearby objects may not be properly aligned, creating a ghosting or echoing effect in the generated panorama, unless correct depth information is taken into account. Taking a cue from the similar problem of dense stereo, we approximate our scene with a Markov random field and use belief propagation to estimate the maximum a posteriori panoramic image for that scene.