Towards More Efficient and Effective LP-based algorithms for MRF Optimization