I think the biggest barrier I've encountered is toolchain support, especially on cross-platform projects. Until a few years ago, it was common to see build notes saying "x platform needs STLport to work because their compiler is borked". Even now, I see issues with people trying to use multiple third-party dependencies tied to different versions of BOOST. This makes linking impossible, meaning you have to go back and rebuild your deps from scratch.
Now that just about everyone has stopped using MSVC++ 6, the STLport mess is behind us. But as soon as TR1 is out the door, we're back to "which versions of which environments support it and get it right" and once again this will slow adoption.
I work on a project begun in C (not C++) in 1992. Deploying modern practices across the legacy codebase would be impossible. Likewise I work on another project that is much closer to the cutting edge of C++ language.
