AV1 Local Warped Motion in Rotating Scenes

This article explains how the AV1 video codec utilizes local warped motion compensation to drastically reduce prediction residuals in rotating scenes. Traditional video compression relies heavily on simple translational motion vectors, which struggle to represent non-linear movements like rotation, zooming, and shearing. By employing affine transformation models at the local block level, AV1 accurately maps rotational trajectories from reference frames, minimizing prediction error and optimizing compression efficiency.

The Limitations of Translational Motion Compensation

Standard motion estimation across legacy codecs assumes that pixels within a coding block move uniformly in horizontal and vertical directions. While this two-dimensional translational vector model works well for linear panning, it fails when applied to rotating objects or rotating camera angles.

In a rotating scene, every pixel within a single block follows a distinct angular trajectory depending on its distance and angle from the rotation center. Forcing a single 2D vector onto a rotating block causes severe structural misalignment along object boundaries. This misalignment produces high prediction residuals—the difference between the original block and the predicted block—requiring a massive number of transform coefficients and bits to correct.

How Local Warped Motion Operates

AV1 solves this issue by incorporating warped motion compensation, which introduces higher-order affine motion models capable of handling rotation, scaling, and shearing. Local warped motion estimates a 2D affine transformation matrix that maps coordinates \((x, y)\) in the current frame to transformed coordinates \((x', y')\) in the reference frame:

\[x' = \alpha x + \beta y + \gamma\] \[y' = \delta x + \epsilon y + \zeta\]

Instead of signaling six full affine parameters for each block—which would incur significant bitrate overhead—AV1 derives these local parameters from the motion vectors of adjacent neighboring blocks. If the neighbor vectors indicate a gradient consistent with rotation or zooming, AV1 constructs an affine model locally.

Reducing Prediction Residuals

Local warped motion reduces prediction residuals in rotating sequences through three primary mechanisms:

  1. Subpixel Grid Transformation: Rather than translating a rigid rectangular block of pixels, the warp model computes a continuous geometric transform across the reference frame. The reference block is warped to mirror the rotational angle of the target block.
  2. High-Precision Interpolation: AV1 applies dedicated 8-tap directional filtering to interpolate pixel values at irregular subpixel positions created by rotation. This avoids the blurring and stair-stepping artifacts common to coarse approximations.
  3. Minimization of High-Frequency Energy: Because the transformed reference patch matches the physical rotation of the scene, structural edges and textures line up precisely with the target block. This eliminates high-frequency edge mismatches, leaving near-zero residual energy across the block.

By accurately reconstructing the geometry of rotating scenes prior to residual calculation, AV1 requires far fewer bits to encode the remaining prediction error, yielding substantially higher visual quality at lower target bitrates.