Partitioning methods

We developed a partitioning scheme, called generalized binary splitting (GBS), that is an extension of previous splitting methods. As in the previous approaches, a video frame is initially divided into so-called coding tree units (CTUs). The CTUs cover squares of N_max×N_max luma samples and are the starting point for a flexible partitioning into smaller blocks as can be seen in the figure above. For the splitting, only binary splits are used. Here, in contrast to previous approaches, binary splits into blocks of unequal size are supported. More precisely, a given rectangle of width W and height H can be split horizontally into two blocks of width W, where the first block has height αH and the second block has height (1-α)H. Here, the split ratio α is chosen out of the set

{ ¹/₂ , ¹/₄ , ³/₄ , ¹/₃ , ²/₃ , ³/₈ , ⁵/₈ , ¹/₅ , ²/₅ , ³/₅ , ⁴/₅ }. 

Vertical splits are supported analogously. The rectangular blocks that result from the partitioning are used for prediction and transform coding of inter prediction residuals. 

The split is coded by a tree as illustrated in the figure above. First, a split flag indicating if a block is further split is transmitted. If the split flag is equal to 1, the split direction and the split ratio are coded. The split direction is signaled as either parallel or perpendicular to the last split. For the first split, i.e., the CTU-level split, perpendicular is defined as a vertical split and parallel as a horizontal split. The split ratio is coded using the binarization shown in the figure. Not all splits in the figure are always possible and two restrictions are imposed in the signalling. First, only splits that result in blocks whose width and height are integral multiples of 4 are allowed. Second, redundancy constraints are imposed in order to guarantee that each final partitioning of a CTU is achievable by only one consecutive sequence of splits.