The above claim that "If you are planning to draw a lot of pixel, it's a lot more efficient to use the image data of the canvas to do pixel drawing" seems to be quite wrong - at least with Chrome 31.0.1650.57 m or depending on your definition of "lot of pixel". I would have preferred to comment directly to the respective post - but unfortunately I don't have enough stackoverflow points yet:
I think that I am drawing "a lot of pixels" and therefore I first followed the respective advice for good measure I later changed my implementation to a simple ctx.fillRect(..) for each drawn point, see http://www.wothke.ch/webgl_orbittrap/Orbittrap.htm
Interestingly it turns out the silly ctx.fillRect() implementation in my example is actually at least twice as fast as the ImageData based double buffering approach.
At least for my scenario it seems that the built-in ctx.getImageData/ctx.putImageData is in fact unbelievably SLOW. (It would be interesting to know the percentage of pixels that need to be touched before an ImageData based approach might take the lead..)
Conclusion: If you need to optimize performance you have to profile YOUR code and act on YOUR findings..