757b505cfd34c64c85ca5b5690ee5293-Supplemental.pdf

We extend the flow-module (introduced in Section 3.2) which aims at reusing contextual information from single-to multi-frame context. We thus consider a context(xi)i J1;cK consisting of c frames. Just as before, for a given levelrk and i in J1;cK, we use F to compute the corresponding fusion maskmki and optical flowfki between intermediate encoded featureseki (from context framexi) and the decoded featuresdks (that we wish to update). We recall that fusion masksmki handle occlusion by indicating for each spatial location the relevance of warped context features e ki = W(eki,fki), that is, whether features e kicorrespond to dks at that location in terms of content, or not.