Generative Partition Networks for Multi-Person Pose Estimation

This paper proposes a new framework, named Generative Partition Network (GPN), for addressing the challenging multi-person pose estimation problem. Different from existing pure top-down and bottom-up solutions, the proposed GPN models the multi-person partition detection as a generative process from joint candidates and infers joint configurations for person instances from each person partition locally, resulting in both low joint detection and joint partition complexities. In particular, GPN designs a generative model based on the Generalized Hough Transform framework to detect person partitions via votes from joint candidates in the Hough space, parameterized by centroids of persons. Such generative model produces joint candidates and their corresponding person partitions by performing only one pass of joint detection. In addition, GPN formulates the inference procedure for joint configurations of human poses as a graph partition problem and optimizes it locally. Inspired by recent success of deep learning techniques for human pose estimation, GPN designs a multi-stage convolutional neural network with feature pyramid branch to jointly learn joint confidence maps and Hough transformation maps. Extensive experiments on two benchmarks demonstrate the efficiency and effectiveness of the proposed GPN.