Depth sweep regression forests for estimating 3d human pose from images

We address the problem of estimating the 3d pose from monocular images. However, instead of learning a regression from image features to the full pose, we regress the positions of the joints in 3d space and infer the pose using a 3d pictorial structure framework. For regression, we rely on regression forests that have been shown to efficiently predict 2d pose from images or 3d pose from depth data. These approaches, however, cannot be directly applied since each local image or depth feature estimates the relative positions of the joints from the feature location. While the relative position is well defined if feature and joint locations are given either in 2d or in 3d, it is, however, not defined if the features are sampled from 2d images without depth information and the joint locations need to be predicted in a 3d world coordinate system.