Layered Motion

The basic idea behind layered motion is to break the image sequence into "layers" which all follow some coherent motion model.

1. Steps

Step 1: Get an approximation of the "local flow" via LK or other methods

Step 2: Obtain a set of initial affine motion hypotheses

• Divide the image into blocks and estimate affine motion parameters in each block by least squares.
• Perform k-means clustering on affine motion parameters

Step 3: Iterate until convergence:

• Assign each pixel to best hypothesis.
  • Pixels with high residual error remain unassigned.
• Perform region filtering to enforce spatial constraints.
• Re-estimate affine motions in each region

The implementation difficulties lie in identifying segments and clustering them appropriately to fit the intended motion models.