Guest Editorial: State of the Art in Image- and Video-Based Human Pose and Motion Estimation

Estimation of articulated human pose from images or video has been a long standing goal of computer vision. Applications for such technology are prevalent across scientific and consumer domains. Despite over a decade of active research, and hundreds of papers, robust solutions to this problem are still unattainable in general unconstrained scenarios. We believe, that progress requires an understanding of the current state of the art and where current methods fail. This necessitates a comprehensive evaluation of current approaches on the same data and with the same metrics. This special pair of issues contain 9 papers that span, and are representative of, recent developments in the field. Among these papers are those that explore algorithmic choices within more traditional methods (e.g., Bayesian Filtering), as well as those that explore new technologies and representations. The topics span kinematic and physicsbased motion priors, generative and discriminative inference methods, part-based models, and parametrized mesh-based body representations. Issues pertaining to the relationship between the number (or speed) of cameras and performance are also explored. To ensure that the methods are directly comparable, we required all authors to report performance on the same datasets (HUMANEVA) and with the same metrics. HUMANEVA is a publicly available dataset initially released for a series of workshops held in conjunction with NIPS (2006) and CVPR (2007). Neither of the two workshops had printed proceedings1but those participating were