Action-Agnostic Human Pose Forecasting

Predicting and forecasting human dynamics is a very interesting butchallenging task with several prospective applications in robotics,health-care, etc. Recently, several methods have been developed for human poseforecasting; however, they often introduce a number of limitations in theirsettings. For instance, previous work either focused only on short-term orlong-term predictions, while sacrificing one or the other. Furthermore, theyincluded the activity labels as part of the training process, and require themat testing time. These limitations confine the usage of pose forecasting modelsfor real-world applications, as often there are no activity-related annotationsfor testing scenarios. In this paper, we propose a new action-agnostic methodfor short- and long-term human pose forecasting. To this end, we propose a newrecurrent neural network for modeling the hierarchical and multi-scalecharacteristics of the human dynamics, denoted by triangular-prism RNN(TP-RNN). Our model captures the latent hierarchical structure embedded intemporal human pose sequences by encoding the temporal dependencies withdifferent time-scales. For evaluation, we run an extensive set of experimentson Human 3.6M and Penn Action datasets and show that our method outperformsbaseline and state-of-the-art methods quantitatively and qualitatively. Codesare available at https://github.com/eddyhkchiu/pose_forecast_wacv/