Tiny Person Pose Estimation via Image and Feature Super Resolution

Although great progress has been achieved on human pose estimation in recent years, we notice the performance drops dramatically when the scale of target person becomes small. In this paper, we start with analysis on tiny person pose estimation and find that the failure is mainly caused by blurriness and ambiguous edges in up-sampled images, which are harmful for pose estimation. Based on the above analysis, we propose to apply an additional super resolution network on top of an existing pose estimation method to better handle tiny persons. Specifically, we propose three super resolution (SR) networks which apply on image level, feature level and both levels, respectively. Furthermore, a novel task-driven loss function tailored to pose estimation is proposed for SR networks. Experimental results on the MPII and MSCOCO datasets show that our proposed pose super resolution methods bring significant improvements over the baseline for tiny persons.