Local vs global energy minimization methods: Application to stereo matching

Energy minimization is often the key point of solving problems in computer vision. For decades, many methods have been proposed (deterministic, stochastic,...). Some can only reach local minimum and others strong local minimum close to the optimal solution (global minimum). Since beginning of 21^th century, minimization based on Graph theory have been generalized to find global minimum of multi-labeling problems. In this work, we study deterministic local minimization methods (Iterative Conditional Modes and Direct Descent Energy), and a stochastic global minimization with an improved Simulated Annealing algorithm. A new approach formulation to help local minimization to converge to a minimum closed to the global one is proposed. This method combines local and global energy constraints in an multiresolution way. We focus on stereo matching application. The improved Simulated Annealing proved to reach global minimum as good as Graph based minimization methods. Promising results of proposed local minimization methods are obtained on Middlebury Stereo database compare to global methods.