Appearance dependent inter-part relationship for human pose estimation

Yumin Suh, Kyoung Mu Lee

Abstract: We propose a new method for human pose estimation from a single image. Since both appearance and locations of different body parts strongly depends on each other in an image, considering their relationship helps identifying the underlying poses. However, most of the existing methods cannot fully utilize this contextual information by using simplified model to make inference tractable. The proposed method models general relationship between body parts based on the convolutional neural networks, while keeping inference tractableble by effectively reducing the search space to a subset of poses by pruning unreliable ones based on the strong unary part detectors. Experimental results demonstrate that the proposed method improves the accuracy than baselines, on FLIC and LSP dataset, while keeping inference and learning tractable.