Interpolated skip mode decision in video compression
A video encoding method determines the best video encoding mode for a macroblock in the SKIP mode and comparing this cost with other modes. This avoids sub-pixel interpolations for fractional pixels. This models the cost function in a quadratic model and computes the cost for a nearest integer pel and plural adjacent integer pels. This permits determination of the coefficients of the quadratic model. An estimated cost is obtained using the actual fractional pel position in the quadratic model. This can be used for obtaining the cost of a P Skip mode. This can be used for the decision of B direct or B skip modes in B frames.
1. A method of video coding comprising:
determining a measure of similarity between a prior frame and a current frame offset for a current macroblock by a motion vector having a fractional pel including
determining a nearest integer pel of said fractional pel,
calculating said measure of similarity for said nearest integer pel,
determining four integer pels adjacent to said nearest integer pel including a first adjacent integer pel immediately left of said nearest integer pel, a second adjacent integer pel immediately right of said nearest integer pel, a third adjacent integer pel immediately above said nearest integer pel and a fourth adjacent integer pel immediately below said nearest integer pel,
calculating said measure of similarity of said nearest integer pel and each of said four adjacent integer pels,
calculating coefficients of a quadratic model of said similarity measure versus pel location using said calculated measure of similarity of said nearest integer pel and said four adjacent integer pels, wherein said quadratic model of said similarity measure versus pel location is:
S ( p )= C 1 x 2 +C 2 x+C 3 y 2 +C 4 y+C 5
where: x and y are spatial coordinate of said pel location; and S(p) is the measure of similarity; and C 1 , C 2 , C 3 , C 4 and C 5 are respective coefficients of the quadratic model of said similarity measure; and said step of calculating coefficients of a quadratic model of said similarity measure versus pel location includes:
C
1
=
S
(
p
1
)
-
2
S
(
p
2
)
-
S
(
p
3
)
2
C
2
=
S
(
p
3
)
-
S
(
p
1
)
2
C
3
=
S
(
p
4
)
-
2
S
(
p
2
)
-
S
(
p
5
)
2
C
4
=
S
(
p
5
)
-
S
(
p
4
)
2
C
5
=
S
(
p
2
)
where: S(p 1 ) is said similarity measure of said first adjacent integer pel; S(p 2 ) is said similarity measure of said nearest integer pel; S(p 3 ) is said similarity measure of said second adjacent integer pel; S(p 4 ) is said similarity measure of said third adjacent integer pel; and S(p 5 ) is said similarity measure of said fourth adjacent integer pel, and
estimating said measure of similarity for said motion vector including a fractional pel using said calculated coefficients of said quadratic model.
2. The method of video encoding of claim 1 , wherein:
said measure of similarity is a Sum of Absolute Differences (SAD) calculated:
S ( p )=Σ| Y ( p )− Y ( c )|
where: p is a pel in a previously reconstructed picture; c is a pel in a current picture; and Y(x) represents a luma value of a relevant pixel.
3. The method of video encoding of claim 1 , wherein:
said measure of similarity is a Sum of Square of Errors (SSE) calculated:
S ( p )=Σ( Y ( p )− Y ( c )) 2
where: p is a pel in a previously reconstructed picture; c is a pel in a current picture; and Y(x) represents a luma value of a relevant pixel.
4. The method of video encoding of claim 1 , further comprising:
calculating a rate distortion cost of said macroblock for a plurality of encoding modes using said estimated measure of similarity;
comparing said calculated rate distortion cost of said macroblock in said plurality of encoding modes;
selecting an encoding mode having a best calculated rate distortion cost; and
encoding said macroblock in said selected mode.
5. The method of video encoding of claim 1 , wherein:
said current macroblock is part of a bi-directional predictive frame (B frame) having forward and backward motion vectors;
said step of determining a nearest integer pel of said fractional pel includes averaging the fractional pels of said forward and backward motion vectors; and
said step of estimating said measure of similarity for said motion vector includes said average a fractional pels.