A Trainable Graph Combination Scheme for Belief Propagation Kai Ju Liu New York University
Images
Pairwise Markov Random Field Basic structure: vertices, edges
Pairwise Markov Random Field Basic structure: vertices, edges Vertex i has set of possible states X i and observed value y i Compatibility between states and observed values, Compatibility between neighboring vertices i and j,
Pairwise MRF: Probabilities Joint probability: Marginal probability: –Advantage: allows average over ambiguous states –Disadvantage: complexity exponential in number of vertices
Belief Propagation
Beliefs replace probabilities: Messages propagate information:
Belief Propagation Example
BP: Questions When can we calculate beliefs exactly? When do beliefs equal probabilities? When is belief propagation efficient? Answer: Singly-Connected Graphs (SCG’s) Graphs without loops Messages terminate at leaf vertices Beliefs equal probabilities Complexity in previous example reduced from 13S 5 to 24S 2
BP on Loopy Graphs Messages do not terminate Energy approximation schemes [Freeman et al.] –Standard belief propagation –Generalized belief propagation Standard belief propagation –Approximates Gibbs free energy of system by Bethe free energy –Iterates, requiring convergence criteria 12 43
BP on Loopy Graphs Tree-based reparameterization [Wainwright] –Reparameterizes distributions on singly-connected graphs –Convergence improved compared to standard belief propagation –Permits calculation of bounds on approximation errors
BP-TwoGraphs Eliminates iteration Utilizes advantages of SCG’s
BP-TwoGraphs Calculate beliefs on each set of SCG’s: – Select set of beliefs with minimum entropy – Consider loopy graph with n vertices Select two sets of SCG’s that approximate the graph –
BP-TwoGraphs on Images Rectangular grid of pixel vertices H i : horizontal graphs G i : vertical graphs horizontal graph vertical graphoriginal graph
Image Segmentation add noise segment
Image Segmentation Results
Image Segmentation Revisited add noise ground truth max-flow ground truth
Image Segmentation: Horizontal Graph Analysis
Image Segmentation: Vertical Graph Analysis
BP-TwoLines Rectangular grid of pixel vertices H i : horizontal lines G i : vertical lines horizontal line vertical lineoriginal graph
Image Segmentation Results II
Image Segmentation Results III
Natural Image Segmentation
Boundary-Based Image Segmentation: Window Vertices Square 2-by-2 window of pixels Each pixel has two states –foreground –background
Boundary-Based Image Segmentation: Overlap
Boundary-Based Image Segmentation: Graph
Real Image Segmentation: Training
Real Image Segmentation: Results
Real Image Segmentation: Gorilla Results
Conclusion BP-TwoGraphs –Accurate and efficient –Extensive use of beliefs –Trainable parameters Future work –Multiple states –Stereo –Image fusion