Multi-Hypothesis based Distributed Video Coding using Error-Correction Decoder Feedback

This paper is dedicated to Distributed Video Coding (DVC) which allows to shift the encoding complexity to the decoder side. We consider DVC architecture, where a side information (SI), which is needed for the error-correction decoder, are selected from multiple hypothesis obtained from set of different traditional and deep learned frame interpolation algorithms. First, we show that the interpolation algorithms could provide different quality of SI depending on a motion model, i.e. the SI averaging proposed in earlier works is not efficient. Therefore, for each bit plane we propose to select the best SI from the available set of hypothesis using feedback of the error-correction decoder: the best hypothesis corresponds to successful error-correction decoding with minimum syndrome bits received from the encoder. Second, in contrast to the previous approaches, we modified the correlation noise model (CNM), so that each bit plane can be processed with the variance obtained from different frame interpolators allowing for more precise noise modeling. Experimental results demonstrate that comparing with DVC with a single frame interpolator this adaptive selection strategy offering significant bitrate savings and gain in peak signal-to-noise ratio (PSNR), particularly in sequences with complex motion and varying group of picture (GOP) lengths.