Boosting neural video codecs by exploiting hierarchical redundancy

Hierarchical Random Access Coding for Deep Neural Video Compression

“…It means with the same quality of reconstructed frames, our model reduces bitrate twice in comparison with the baseline SSF model. In addition, we also compare to two previous methods, B-EPIC [33], and BoostSSF [15] which also use SSF as a baseline model for I-frame and Pframe. From the results shown in Table 2, among SSF-based methods, ours are the best and outperforms the others in terms of bitrate saving for PSNR with respect to the anchor SSF model.…”

“…Deep neural networks (DNNs) based video compression has attracted extensive attention from improving existing coding tools [3], [5], [8], [9], [31] to end-to-end neural video coding [4], [6], [7], [10], [11], [12], [13], [14], [15], [16]. In [9] proposed to use a hybrid spatial-temporal entropy modeling for video compression.…”

“…In the other approach, Hu et al [6] applied coarse-to-fine coding with hyperprior-guided mode prediction, and Agustsson et al [16] proposed scale-space flow instead of an explicit motion estimation network for end-to-end video compression. Pourreza et al [15] improves the coding efficiency of a neural video compression model by exploiting hierarchical redundancy between frames for both residual and flow information. Sun et al [34] proposed a spatiotemporal entropy model for learned video compression.…”

“…In all-intra coding configuration, each frame is encoded using a neural image compression model separately, it does not exploit the temporal redundancy between frames that are the most powerful coding tool in classical video coding with the highest contribution to coding efficiency of the classical video codecs, thanks to temporal inter-frame prediction. For low delay configuration, several state-of-the-art neural video compression models [11], [12], [13], [14], [15], [16] exploit inter-frame prediction with motion estimation and motion compensation networks. However, they just apply inter-frame prediction at one side of the current coding frame, the previous frame, or P-frame prediction, due to the low delay requirement of encoding processing.…”