Siamese Networks Based People Tracking for 360-degree Videos with Equi-angular Cubemap Format

This paper proposes a deep learning based pedestrian tracking scheme for 360-degree videos using equiangular cubemap (EAC) format. To be robust against content discontinuity of EAC images, this paper proposes an efficient face stitching scheme such that the tracker keeps tracking across adjacent faces and avoids raising geometric deformation simultaneously. By referring to statistics of score maps from efficient fully-convolutional siamese networks, the proposed mechanism of template update determines the timing of update. Experimental results show that the proposed tracker operates at 60 fps and outperforms the fully convolutional siamese networks based tracker on 360-degree videos with EAC format both in precision plots and success plots.