Despite of the recent success of neural networks for human pose estimation,current approaches are limited to pose estimation of a single person and cannothandle humans in groups or crowds. In this work, we propose a method thatestimates the poses of multiple persons in an image in which a person can beoccluded by another person or might be truncated. To this end, we considermulti-person pose estimation as a joint-to-person association problem. Weconstruct a fully connected graph from a set of detected joint candidates in animage and resolve the joint-to-person association and outlier detection usinginteger linear programming. Since solving joint-to-person association jointlyfor all persons in an image is an NP-hard problem and even approximations areexpensive, we solve the problem locally for each person. On the challengingMPII Human Pose Dataset for multiple persons, our approach achieves theaccuracy of a state-of-the-art method, but it is 6,000 to 19,000 times faster.
展开▼
