Exploiting Long-Term Redundancies in Reconstructed Video

In this paper we propose an enhancement to HEVC that allows the usage of homographies for motion compensation between temporally far distant frames. First, we show the benefit of using homographies compared to block-based motion compensation, particularly for long-term motion compensation in the presence of zoom. Then we demonstrate that beyond the case where content is temporally repeated, long-term global motion compensation leads to a better compression ratio in the presence of zooming out. We show that this is due to the change in scale on one hand, and the effects of compression on the reference pictures on the other. Using a model for the distortion of images caused by compression, it will be shown that a finer scale reconstructed image that is warped into the perspective of the image to be encoded can constitute a better predictor than an immediately neighboring picture. We compared the HM software with an enhanced system using long-term global motion compensation with homographies and found that the enhanced system achieves Bjøntegaard Delta Rate savings of up to 38%. With the mechanism behind the gains unveiled, we finally try to find an optimal scale ratio between the frame to be encoded and the reference frame.