We present a novel approach for multi-person pose estimation (MPPE) using implicit column generation and nested benders decomposition. We formulate MPPE as a set packing problem over the set of person hypothesis (poses) in an image where the set of poses is the power set of detections of body parts in the image. We model the quality of a pose as a function of its members as described by a tree structured deformable part model.
Since we cannot enumerate the set of poses we attack inference using implicit column generation where the pricing problem is structured as a dynamic program and dual optimal inequalities are easily computed. We exploit structure in the dynamic program to permit efficient inference using nested Benders decomposition. We demonstrate the effectiveness of our approach on the MPII human pose annotation benchmark data set.
Bienvenue à tous!