Nonstationary Shape Activities: Dynamic Models for Landmark Shape Change and Applications

摘要: Our goal is to develop statistical models for the shape change of a configuration ?landmark? points (key interest) over time and use these filtering tracking automatically extract landmarks, synthesis, detection. The term ?shape activity? was introduced in recent work denote particular stochastic model dynamics landmark shapes (dynamics after global translation, scale, rotation effects are normalized for). In that work, only stationary sequences were proposed. But most ?activities? set e.g., running, jumping, or crawling, have large changes with respect initial hence nonstationary. key contribution this novel approach define generative both 2D 3D nonstationary sequences. Greatly improved performance using proposed demonstrated sequentially noise-corrupted configurations compute Minimum Mean Procrustes Square Error (MMPSE) estimates true human activity videos, i.e., predict locations landmarks (body parts) prediction faster more accurate extraction from current image.