ME: Deformable Block Matching (Warping)
PROS: works with complex motions (translation, rotation, scaling)
CONS: doesn’t try to simultaneously classify the pixels
select grid vertices in the previous frame
- such mesh partitions the frame into non-overlapping quadrilateral blocks
- grid points should reside on distinct image features (edges or corners)
- start with uniform grid, then optimize by locally searching for edges
find matching grid vertices in the current frame
- take square block centered around each vertex and do block matching
- when computing MAD, weight its terms increasingly towards the center
calculate grid vertices in the interpolated frame
- assume linear motion of grid vertices
compute forward and backward warped frames
- divide each quadrilateral into two triangles
- use affine mapping of triangles to find an appropriate triangle
take distance-weighted average of gray values for each pixel
- resulting value is not an integer, so use linear interpolation