Here's an explanation of the Photoshop gag: http://www.threadless.com/profile/433934/elleevee/blog/493381/Threadless
Basically they're using the displace filter to shove only the red channel or the blue+green channels of an image left or right. The heightmap just attenuates the distance that you displace each pixel. I believe the displace filter interpolates the inbetween values, i.e. if the pixel at (10,0) was displaced to the left by 3 pixels, and pixel (11,0) was displaced to the left by only 1, then the two pixels in between the target pixels would be interpolated from the two original values at 66% and 33%.
That interpolation effectively is covering for missing data: if you really had two viewpoints of your scene, then these pixels would represent information hidden in the single view. I can imagine an upgrade to the method outlined above, where reconstruction algorithms similar to Photoshop's content aware fill could take a better stab at filling in the missing information.