I have the exact same use case and just tried both. Loading a lot of static sprites (from the same image) is faster in three.js for 5000 sprites and above, but animating only a few of those sprites give better framerates in pixi (again, for 5000 sprites). (This was tested on Chrome and IE9 on desktop)
The biggest difference was with the Canvas renderer, where pixi's autodetect gives the same results as the WebGL (if slower) for the same code, but three.js's Canvas renderer doesn't support the Sprite type meaning to achieve portable code you have to use Particles. If you don't use sprites all that much and mostly have quads or triangles, that wouldn't be an issue.