Uniting holistic and part-based attitudes for accurate and robust deep human pose estimation

摘要

Deep learning has been utilized in many intelligent systems, including computer vision techniques. Human pose estimation is one of the popular tasks in computer vision that has benefited from modern feature learning strategies. In this regard, recent advances propose part-based approaches since pose estimation based on parts can produce more accurate results than when the human shape is considered holistically as one unbreakable, but deformable object. However, in real-word scenarios, problems like occlusion and cluttered background make difficulties in part-based methods. In this paper, we propose to unite the two attitudes of the part-based and the holistic pose predictions to make more accurate and more robust estimations. These two schemes are modeled using convolutional neural networks as regression and classification tasks in order, and are combined in three frameworks: multitasking, series, and parallel. Each of these settings has its own advantages, and the experimental results on the LSP test set demonstrate that it is essential to observe subjects, both based on parts and holistically in order to achieve more accurate and more robust estimation of human pose in challenging scenarios.