In this work, we address the problem of human pose estimation in still images by proposing a holistic model for learning the appearance of the human body from image patches. These patches, which are randomly chosen, are used for extracting features and training a regression forest. During training, a mapping between image features and human poses, defined by joint offsets, is learned; while during prediction, the body joints are estimated with an efficient mode-seeking algorithm. In comparison to other holistic approaches, we can recover body poses from occlusion or noisy data. We demonstrate the power of our method in two publicly available datasets and propose a third one. Finally, we achieve state-of-the-art results in comparison to other approaches.
展开▼
