It's fairly disingenuous to say that implementing polygon-based (the article calls it "vectorial") collision detection on platforms is "very difficult to implem

It's fairly disingenuous to say that implementing polygon-based (the article calls it "vectorial") collision detection on platforms is "very difficult to implement properly" without a physics engine. The collision detection code for this can be done in probably less than a thousand lines (probably less if everything is OBBs). Adding support for arcs, and even bezier curves becomes much, much easier as well (for both, probably another thousand lines, at most). It is also easier to add support for continuous collision detection this way, should that be something you want (although this would increase the amount of code a fair amount). FWIW, I used to go the pixel-based route, but after having done it this way I'm not sure I would do it again. It ends up being much cleaner, more robust, and easier to extend. I'd go as far as to say that your collision detection and physics code shouldn't have any knowledge of pixels, only your renderer (and maybe the gameplay code, if you insist) should have an idea about these. I will say that it requires a stronger math background, however, but in all honesty, probably not that much stronger than is required to make the rest of a game.