Sameer Agarwal and I just uploaded a paper to arχiv.org: Continuous Optimization for Fields of Experts Denoising Works. We show that simple non-linear least-squares is better than sophisticated discrete optimization methods for certain image denoising problems.
I have used these denoising problems as a bechmark and motivation when developing generalized roof duality. But as it turns out, continuous optimization is much better. 🙂
Update: I should add that the source code used for the experiments in available in Ceres; see https://code.google.com/p/ceres-solver/ .