We introduce a novel data-driven mean-shift belief propagation (DDMSBP) method for non-Gaussian MRFs, which often arise in computer vision applications. With the aid of scale space theory, optimization of non-Gaussian, multimodal MRF models using DDMSBP becomes less sensitive to local maxima. This is a significant improvement over standard BP inference, and extends the range of methods that are computationally tractable. In particular, when pair-wise potentials are Gaussians, the time complexity of DDMSBP becomes bilinear in the numbers of states and nodes in the MRF. Experimental results from simulation and non-rigid deformable neuroimage registration demonstrate that our method is faster and more accurate than state-of-the-art inference algorithms.

Symmetry is one of the important cues for human and machine perception of the world. For over three decades, automatic symmetry detection from images/patterns has been a standing topic in computer vision. We present a timely, systematic, and quantitative performance evaluation of three state of the art discrete symmetry detection algorithms. This evaluation scheme includes a set of carefully chosen synthetic and real images presenting justified, unambiguous single or multiple dominant symmetries, and a pair of well-defined success rates for validation. We make our 176 test images with associated hand-labeled ground truth publicly available with this paper. In addition, we explore the potential contribution of symmetry detection for object recognition by testing the symmetry detection algorithm on three publicly available object recognition image sets (PASCAL VOC'07, MSRC and Caltech-256). Our results indicate that even after several decades of effort, symmetry detection in real-world images remains a challenging, unsolved problem in computer vision. Meanwhile, we illustrate its future potential in object recognition.