Abstract

A scale, rotation and articulation invariant method is pro- posed to match human sub jects in images. Different from the widely used pictorial structure scheme, the proposed method directly matches body parts to image regions which are obtained from ob ject independent proposals and successively merged superpixels. Body part region match- ing is formulated as a graph matching problem. We globally assign a body part candidate to each node on the model graph so that the over- all configuration satisfies the spatial layout of a human body plan, part regions have small overlap, and the part coverage follows proper area ratios. The proposed graph model is non-tree and contains high order hyper-edges. We propose an efficient method that finds global optimal solution to the matching problem with a sequence of branch and bound procedures. The experiments show that the proposed method is able to handle arbitrary scale, rotation, articulation and match human sub jects in cluttered images.