Preserving Modes and Messages via Diverse Particle Selection
Jason Pacheco, Silvia Zuffi, Michael J. Black, Erik B. Sudderth · 2014
In applications of graphical models arising in do-mains such as computer vision and signal pro-cessing, we often seek the most likely config-urations of high-dimensional, continuous vari-ables. We develop a particle-based max-product algorithm which maintains a diverse set of pos-terior mode hypotheses, and is robust to initial-ization. At each iteration, the set of hypothe-ses at each node is augmented via stochastic pro-posals, and then reduced via an efficient selec-tion algorithm. The integer program underlying our optimization-based particle selection mini-mizes errors in subsequent max-product mes-sage updates. This objective automatically en-courages diversity in the maintained hypotheses, without requiring tuning of application-specific distances among hypotheses. By avoiding the stochastic resampling steps underlying particle sum-product algorithms, we also avoid common degeneracies where particles collapse onto a sin-gle hypothesis. Our approach significantly out-performs previous particle-based algorithms in experiments focusing on the estimation of human pose from single images. 1.