Fast and Accurate Pose Estimation in Videos based on Knowledge Distillation and Pose Propagation

Abstract

Existing video-based human pose estimation methods typically adopt large networks to perform body joints localization on all frames. Despite of impressive accuracy performance, the relatively high memory and computation requirements significantly burden their applicability on resource-constraint systems (e.g., embedded devices). To solve this issue, this paper proposes a novel yet effective lightweight framework, called FVPE, for fast and accurate human pose estimation in videos. Specifically, FVPE adopts the knowledge distillation (KD) strategy to train a small pose estimator network, which is capable of executing rapidly with low computational cost. To increase the overall efficiency, FVPE exploits the temporal coherence between successive video frames and explicitly propagates body joints from previous frames rather than naively extracting them using a pose estimator. Furthermore, FVPE introduces an online key-frame selection scheme to decide whether the current pose should be calculated by the pose estimator or be propagated from the previous key-frame, being able to flexibly deal with video sequences with different length, frame rate, pose complexity, etc.. Experiments on Penn Action and Sub-JHMDB datasets demonstrate that the proposed method achieves comparative accuracy, but with substantial speed-up.