Machine Learning Based Efficient QT-MTT Partitioning Scheme for VVC Intra Encoders

Light-Weight CNN-Based VVC Inter Partitioning Acceleration

“…Consequently, CUs at a specific depth in the tree do not correspond to the same size and shape, introducing dependence between the MT splits along the partition path. This dependence partly explains the decrease in partition prediction accuracy as the depth of the partitioning tree increases, as observed in recent studies [15,24] presented in the previous section.…”

“…This method is fast in the sense that a single vector is computed for each CTU. Nevertheless, it is observed in [24], that the predictions are more accurate at higher levels of the partitioning tree. Hence, they propose improving the decisions by adding 16 trained DTs to process the CNN output, introducing additional complexity to the method.…”

“…A complete partitioning acceleration of inter coding by CNN is proposed in [15]. A vector containing probabilities of the existence of split boundaries in the partition is predicted similarly to [24]. This method is fast in the sense that a single vector is computed for each CTU.…”

“…Previous partition representations at CTU level have typically used binary vectors to depict split boundaries. In [14,15,24], the authors intend to predict the split boundary of each 4x4 sub-block in CTU. Lately, Wu et al improve this representation in [16] by proposing hierarchical boundaries.…”

“…In the case of [14], conversions are carried out by computing the average probability at the location of the specific split. [15] and [24] convert boundary vectors to split decisions by DT models separately trained for different CU sizes. Secondly, different partition paths could be deduced from a particular partition presented in a geometric way.…”