Motion Partition

Object-Boundary-Based motion partition scheme

（e）is the best approximate of foreground picture but wastes coding. (d) is a trade off.

Partition by two component with shape. After partition map is obtained, motion estimation is performed independently for each partition, generating two motion vectors for the current macroblock. Then merge them.

OBJECT BOUNDARY BASED MOTION PARTITION FOR VIDEO CODING

Wedge Partition

0 and r are deciding parameter.

Since motion compensation using wedge partitioning is expected to work better for blocks that fall along boundaries of moving objects, it is reasonable to assume that preselecting partitions based on the image edges may also achieve satisfactory results. In particular, within an NxN block, we first use an edge detection algorithm such as Canny or Sobel, followed by a thresholding operation. This yields a two-dimensional logical array of the same size as the block indicating edges. Next, the pixels with positive response are mapped into Euclidian pixel coordinates yielding (x , y) pairs, and a linear regression (based on least square error) is applied to these points. The result is a general equation of a line: y = ax + b, or x = c (when the regression is not a function). The line is converted to polar coordinates and is associated to a wedge partition. Let (x0 , YA) be the block center, N be the block size and 00 be an offset that depends on the sign of axO + yo+ b, such that
00 = 0° when axO+ yo+ b is positive and 00 = 180°

ON MACROBLOCK PARTITION FOR MOTION COMPENSATION