The author neglects to point out that every image will waste a http request downloading the svg file so this is going to give poorer performance than having the logic in your css (perhaps he considered it too obvious).
But, the benefit of this technique (as far as I can tell) is decoupling, and any kind of embedding re-couples the data. Maybe in an acceptable way, if you can do it at a deploy step. It depends on what the use case is.
The method at this link: http://estelle.github.io/clowncar/bgonly.html only does two http requests, the svg and the image used. We can bring that down to 1 by using data URI for the svg logic only.