Articulated human pose estimation in natural images

In this thesis the problem of estimating the 2-D articulated pose, or configuration of a person in unconstrained images such as consumer photographs is addressed. Contributions are split among three major chapters. In previous work the Pictorial Structure Model approach has proven particularly successful, and is appealing because of its moderate computational cost. However, the accuracy of resulting pose estimates has been limited by the use of simple representations of limb appearance. In this thesis strong discriminatively trained limb detectors combining gradient and colour segmentation cues are proposed. The approach improves significantly on the “iterative image parsing” method which was the state-of-the-art at the time, and shows significant promise for combination with other models of pose and appearance. In the second part of this thesis higher fidelity models of pose and appearance are proposed. The aim is to tackle extremely challenging properties of the human pose estimation task arising from variation in pose, anatomy, clothing, and imaging conditions. Current methods use simple models of body part appearance and plausible configurations due to limitations of available training data and constraints on computational expense. It is shown that such models severely limit accuracy. A new annotated database of challenging consumer images is introduced, an order of magnitude larger than currently available datasets. This larger amount of data allows partitioning of the pose space and the learning of multiple, clustered Pictorial Structure Models. A relative improvement in accuracy of over 50% is achieved compared to the standard, single model approach. In the final part of this thesis the clustered Pictorial Structure Model framework is extended to handle much larger quantities of training data. Furthermore it is shown how to utilise Amazon Mechanical Turk and a latent annotation update scheme to achieve high quality annotations at low cost. A significant increase in pose estimation accuracy is presented, while the computational expense of the framework is improved by a factor of 10.