How AV1 Uses Global Motion Estimation

The AV1 video codec leverages Global Motion Estimation (GME) to detect and compensate for whole-frame movement, such as camera panning, tilting, rotation, and zooming, across adjacent video frames. Instead of encoding redundant motion vectors for hundreds of individual blocks, AV1 calculates a unified mathematical model that describes the background movement of the scene. By signaling these global transformation parameters once at the frame level, AV1 allows individual blocks to inherit the camera trajectory, drastically cutting down motion vector data and improving overall compression efficiency.

The Purpose of Global Motion Estimation

In conventional video coding, motion compensation divides a frame into blocks and calculates a motion vector for each block relative to a reference frame. When a camera moves across a scene, nearly every background block shifts in a uniform or mathematically predictable direction. Transmitting individual motion vectors for all of these blocks wastes significant bitrate.

AV1 solves this by using GME to model the global movement caused by camera motion across adjacent frames, separating the background movement from independently moving foreground objects.

Transformation Models in AV1

AV1 implements four mathematical models of increasing complexity to capture various types of camera trajectories between frames:

  1. Identity: Represents no global motion between the reference and the current frame.
  2. Translation: Uses two parameters to represent simple two-dimensional horizontal and vertical shifts (panning and tilting).
  3. Affine: Uses six parameters to capture translation, rotation, isotropic zooming, and non-isotropic stretching (shear).
  4. Projective (Perspective): Uses eight parameters to model full perspective shifts, such as when a camera tilts at an angle relative to a flat plane, causing 3D perspective distortion.

By supporting affine and projective models, AV1 can accurately track complex physical camera movements that older codecs could only approximate with local block translations.

How AV1 Computes Global Motion

To determine the global motion parameters across adjacent frames, the AV1 encoder follows a multi-step estimation pipeline:

  1. Feature Detection and Matching: The encoder detects salient keypoints across adjacent frames or relies on pre-calculated block-level motion vectors to find corresponding point pairs between the current frame and the reference frame.
  2. Outlier Rejection (RANSAC): Real-world scenes often feature moving foreground objects (e.g., people, vehicles) moving independently of the camera. AV1 uses algorithms like Random Sample Consensus (RANSAC) to discard these foreground vectors and isolate the dominant, coherent motion of the background.
  3. Parameter Optimization: The remaining inlier points are fitted into the translation, affine, or projective equations. The encoder refines the transformation matrices to minimize the prediction error over the entire frame.
  4. Rate-Distortion Optimization (RDO): The encoder evaluates the cost of transmitting the model against the quality gain it provides. If a higher-order model (like projective) does not provide enough coding gain over a simpler model (like translation or identity), the simpler model is chosen to save bitstream overhead.

Signaling and Block-Level Application

Once calculated, the global motion parameters are encoded in the frame header for each active reference frame.

During block-level prediction, the encoder can set a block's motion mode to GLOBALMV. When this mode is flagged, the decoder derives the block’s motion vector directly from the frame's global transformation matrix rather than reading unique motion vector coordinates from the bitstream. If a block belongs to an independently moving object, the encoder bypasses the global model and transmits standard local motion vectors.

Efficiency Gains

By delegating predictable camera paths to global mathematical transforms, AV1 eliminates massive amounts of motion vector overhead in sequences with sweeping camera sweeps, aerial drone footage, and sports broadcasts. The bits saved from motion signaling are directly reallocated to preserving fine textures and high-frequency details.