Non-Square Partition Shapes in AV1 Codec

The AOMedia Video 1 (AV1) specification enhances coding efficiency by introducing a flexible block partitioning structure that departs from traditional purely square quad-tree splits. This article details the non-square partition shapes implemented in AV1, including standard 2:1 and 1:2 rectangles, extended 4:1 and 1:4 partitions, and asymmetrical "T-shaped" partitions, explaining how these geometric options enable the codec to better match natural image features and motion.

The Evolution of Block Partitioning in AV1

In modern video coding, frames are divided into large units called Superblocks (typically 128x128 or 64x64 pixels in AV1). Earlier standards like VP9 and H.264 heavily favored symmetric square blocks, with limited rectangular support.

AV1 expands on this by offering a 10-way partition tree. While square partitions (PARTITION_NONE and PARTITION_SPLIT) remain the baseline, AV1 introduces several non-square shapes to better isolate directional edges, text, and moving boundaries.

2:1 and 1:2 Rectangular Partitions

AV1 includes fundamental two-way rectangular divisions:

• Horizontal Split (PARTITION_HORZ): Divides a square block horizontally into two equal rectangular blocks, each with a 2:1 aspect ratio (width is twice the height).
• Vertical Split (PARTITION_VERT): Divides a square block vertically into two equal rectangular blocks, each with a 1:2 aspect ratio (height is twice the width).

These blocks can often be further recursively partitioned down to sizes as small as 4x8 or 8x4 pixels.

4:1 and 1:4 Extended Aspect Ratio Partitions

To handle fine horizontal or vertical structures—such as scrolling tickers, sharp horizons, or vertical stripes—AV1 introduces 4:1 and 1:4 aspect ratio partitions:

• Horizontal 4:1 (PARTITION_HORZ_4): Splits a parent block horizontally into four equal strips, each with a 4:1 width-to-height ratio.
• Vertical 1:4 (PARTITION_VERT_4): Splits a parent block vertically into four equal strips, each with a 1:4 width-to-height ratio.

These extended shapes allow the encoder to isolate narrow, continuous directional details without needlessly splitting the surrounding uniform areas into tiny square blocks.

Asymmetrical "T-Shaped" Partitions

AV1 adds four non-symmetric partition types, often referred to as "T-type" or "AB-type" partitions. These modes divide a square block into three sub-blocks: one half-sized rectangle and two quarter-sized squares or rectangles.

• PARTITION_HORZ_A: The top half is split vertically into two blocks, while the bottom half remains an undivided 2:1 rectangle.
• PARTITION_HORZ_B: The top half remains an undivided 2:1 rectangle, while the bottom half is split vertically into two blocks.
• PARTITION_VERT_A: The left half is split horizontally into two blocks, while the right half remains an undivided 1:2 rectangle.
• PARTITION_VERT_B: The left half remains an undivided 1:2 rectangle, while the right half is split horizontally into two blocks.

These shapes accommodate image regions where detail or motion occurs predominantly on one side of a block, eliminating the requirement to split the entire area uniformly.

Coding Benefits of Non-Square Shapes

The addition of these non-square partitions provides clear benefits:

• Reduced Overhead: Encoders can cover irregular image edges with fewer total transform blocks, significantly lowering signaling overhead.
• Better Prediction Accuracy: Intra prediction and inter-frame motion compensation perform better when block boundaries align closely with natural object boundaries.
• Optimal Transform Sizing: Non-square partitions map directly to AV1's extended library of 2:1, 1:2, 4:1, and 1:4 2D discrete cosine transforms (DCT) and asymmetric discrete sine transforms (ADST).