I can't give you the best practice, but I'm in favor of use of a camera as opposed to moving the object itself.

Here's why: a project of mine started with using the translate() method.  It made sense to me at the time because I was handling only one object, and it was easy to code.

I kind of regretted this later because I found myself stuck with the approach.  I realized it was a mistake when I needed to create a series of labels that needed to face the user's viewpoint, and I had to keep track of the main object's rotation in order to do it, which led to a very clunky implementation.  So as the models being translated became more complicated, use of camera made more and more sense.  It works just fine now, but I had to rebuild a system just to make it behave as if the camera was moving, which isn't a very smart design.

Hope this helps, and I'd like to know others' opinions as well.