Difference between revisions of "H.264 Prediction"

From MultimediaWiki
Jump to: navigation, search
(16x16 Prediction Modes: remove predictors from tables if they don't figure into the calculations)
m (Plane: ohh, Mike, Mike...)
 
Line 476: Line 476:
  
 
   a = 16 * (L15 + T15 + 1) - 7*(V+H)
 
   a = 16 * (L15 + T15 + 1) - 7*(V+H)
   for (j = 0..1Proxy-ConProxy-Connection: keep-alive
+
   for (j = 0..15)
CacProProxy-Proxy-Connection: keep-alive
+
Cache-Control: max-age=0
+
 
+
nnection: keep-alive
+
Cache-Control: max-age=0
+
 
+
-Connection: keep-alive
+
Cache-Control: max-age=0
+
 
+
-Control: max-age=0
+
 
+
ction: keep-alive
+
Cache-Control: max-age=0
+
 
+
 
+
 
     for (i = 0..15)
 
     for (i = 0..15)
 
       b = a + V * j + H * i
 
       b = a + V * j + H * i

Latest revision as of 02:50, 23 April 2011

This page documents the various prediction methods used in H.264 and related formats such as Sorenson Video 3, RealVideo 4, and On2 VP8.

4x4 Prediction Modes

4x4 prediction modes vary between different codecs. While they are almost the same for H.264 and Sorenson Video 3, RealVideo 4 has a different order for these modes and some of them significantly differ from H.264 counterparts (by using left predictors where H.264 does not and down left predictors which are not used elsewhere).

Vertical

  • H.264: mode 0
  • SVQ3: mode 0
  • RV40: mode 1
  • VP8: not used
    | T0  T1  T2  T3
---------------------
    | T0  T1  T2  T3
    | T0  T1  T2  T3
    | T0  T1  T2  T3
    | T0  T1  T2  T3

Vertical (VP8)

  • H.264: not used
  • SVQ3: not used
  • RV40: not used
  • VP8: mode 2
 LT | T0  T1  T2  T3  T4
------------------------
    |  a   b   c   d
    |  a   b   c   d
    |  a   b   c   d
    |  a   b   c   d

where:

 a = (LT + 2*T0 + T1 + 2) >> 2
 b = (T0 + 2*T1 + T2 + 2) >> 2
 c = (T1 + 2*T2 + T3 + 2) >> 2
 d = (T2 + 2*T3 + T4 + 2) >> 2

Horizontal

  • H.264: mode 1
  • SVQ3: mode 1
  • RV40: mode 2
  • VP8: not used
    | 
---------------------
 L0 | L0  L0  L0  L0
 L1 | L1  L1  L1  L1
 L2 | L2  L2  L2  L2
 L3 | L3  L3  L3  L3

Horizontal (VP8)

  • H.264: not used
  • SVQ3: not used
  • RV40: not used
  • VP8: mode 3
 LT | 
--------------------
 L0 |  a   a   a   a
 L1 |  b   b   b   b
 L2 |  c   c   c   c
 L3 |  d   d   d   d

where:

 a = (LT + 2*L0 + L1 + 2) >> 2
 b = (L0 + 2*L1 + L2 + 2) >> 2
 c = (L1 + 2*L2 + L3 + 2) >> 2
 d = (L2 + 2*L3 + L3 + 2) >> 2

DC

  • H.264: mode 2
  • SVQ3: mode 2
  • RV40: mode 0
  • VP8: mode 0
    | T0  T1  T2  T3
---------------------
 L0 |  a   a   a   a
 L1 |  a   a   a   a
 L2 |  a   a   a   a
 L3 |  a   a   a   a

where:

if top and left predictors are available
  a = (T0 + T1 + T2 + T3 + L0 + L1 + L2 + L3 + 4) / 8
else if top predictors are available
  a = (T0 + T1 + T2 + T3 + 2) / 4
else if left predictors are available
  a = (L0 + L1 + L2 + L3 + 2) / 4
else
  a = 128

Note that the VP8 reference code does not make any provisions for either or both sets of predictors to be missing.

Diagonal Down/Left

  • H.264: mode 3
  • SVQ3: not used
  • RV40: not used
  • VP8: mode 4
    | T0  T1  T2  T3  T4  T5  T6  T7
-------------------------------------
    |  a   b   c   d
    |  b   c   d   e
    |  c   d   e   f
    |  d   e   f   g

where:

 a = (T0 + 2*T1 + T2 + 2) / 4
 b = (T1 + 2*T2 + T3 + 2) / 4
 c = (T2 + 2*T3 + T4 + 2) / 4
 d = (T3 + 2*T4 + T5 + 2) / 4
 e = (T4 + 2*T5 + T6 + 2) / 4
 f = (T5 + 2*T6 + T7 + 2) / 4
 g = (T6 + 3*T7      + 2) / 4

Diagonal Down/Left (SVQ3)

  • H.264: not used
  • SVQ3: mode 3
  • RV40: not used
  • VP8: not used
    |     T1  T2  T3
---------------------
    |  a   b   c   c
 L1 |  b   c   c   c
 L2 |  c   c   c   c
 L3 |  c   c   c   c

where:

 a = (L1 + T1) / 2
 b = (L2 + T2) / 2
 c = (L3 + T3) / 2

Diagonal Down/Left (RV40)

  • H.264: not used
  • SVQ3: not used
  • RV40: mode 4
  • VP8: not used
    | T0  T1  T2  T3  T4  T5  T6  T7
-------------------------------------
 L0 |  a   b   c   d
 L1 |  b   c   d   e
 L2 |  c   d   e   f
 L3 |  d   e   f   g
 L4 |
 L5 |
 L6 |
 L7 |

where:

 a = (T0 + 2*T1 + T2 + L0 + 2*L1 + L2 + 4) / 8
 b = (T1 + 2*T2 + T3 + L1 + 2*L2 + L3 + 4) / 8
 c = (T2 + 2*T3 + T4 + L2 + 2*L3 + L4 + 4) / 8
 d = (T3 + 2*T4 + T5 + L3 + 2*L4 + L5 + 4) / 8
 e = (T4 + 2*T5 + T6 + L4 + 2*L5 + L6 + 4) / 8
 f = (T5 + 2*T6 + T7 + L5 + 2*L6 + L7 + 4) / 8
 g = (T6 +   T7      + L6 +   L7      + 2) / 4

Diagonal Down/Right

  • H.264: mode 4
  • SVQ3: mode 4
  • RV40: mode 3
  • VP8: mode 5
 LT | T0  T1  T2  T3
---------------------
 L0 |  d   e   f   g
 L1 |  c   d   e   f
 L2 |  b   c   d   e
 L3 |  a   b   c   d

where:

 a = (L3 + 2*L2 + L1 + 2) / 4
 b = (L2 + 2*L1 + L0 + 2) / 4
 c = (L1 + 2*L0 + LT + 2) / 4
 d = (L0 + 2*LT + T0 + 2) / 4
 e = (LT + 2*T0 + T1 + 2) / 4
 f = (T0 + 2*T1 + T2 + 2) / 4
 g = (T1 + 2*T2 + T3 + 2) / 4

Vertical/Right

  • H.264: mode 5
  • SVQ3: mode 5
  • RV40: mode 5
  • VP8: mode 6
 LT | T0  T1  T2  T3
---------------------
 L0 |  a   b   c   d
 L1 |  e   f   g   h
 L2 |  i   a   b   c
    |  j   e   f   g

where:

 a = (LT + T0 + 1) / 2
 b = (T0 + T1 + 1) / 2
 c = (T1 + T2 + 1) / 2
 d = (T2 + T3 + 1) / 2
 e = (L0 + 2*LT + T0 + 2) / 4
 f = (LT + 2*T0 + T1 + 2) / 4
 g = (T0 + 2*T1 + T2 + 2) / 4
 h = (T1 + 2*T2 + T3 + 2) / 4
 i = (LT + 2*L0 + L1 + 2) / 4
 j = (L0 + 2*L1 + L2 + 2) / 4

Horizontal/Down

  • H.264: mode 6
  • SVQ3: mode 6
  • RV40: mode 8
  • VP8: mode 8
 LT | T0  T1  T2  
---------------------
 L0 |  a   b   c   d
 L1 |  e   f   a   b
 L2 |  g   h   e   f
 L3 |  i   j   g   h

where:

 a = (LT + L0 + 1) / 2
 b = (L0 + 2*LT + T0 + 2) / 4
 c = (LT + 2*T0 + T1 + 2) / 4
 d = (T0 + 2*T1 + T2 + 2) / 4
 e = (L0 + L1 + 1) / 2
 f = (LT + 2*L0 + L1 + 2) / 4
 g = (L1 + L2 + 1) / 2
 h = (L0 + 2*L1 + L2 + 2) / 4
 g = (L2 + L3 + 1) / 2
 j = (L1 + 2*L2 + L3 + 2) / 4

Vertical/Left

  • H.264: mode 7
  • SVQ3: mode 7
  • RV40: mode 6
  • VP8: mode 7
    | T0  T1  T2  T3  T4  T5  T6 
---------------------------------
    |  a   b   c   d
 L1 |  f   g   h   i
 L2 |  b   c   d   e
 L3 |  g   h   i   j
 L4 |

where:

 a = (T0 + T1 + 1) / 2
 b = (T1 + T2 + 1) / 2
 c = (T2 + T3 + 1) / 2
 d = (T3 + T4 + 1) / 2
 e = (T4 + T5 + 1) / 2
 f = (T0 + 2*T1 + T2 + 2) / 4
 g = (T1 + 2*T2 + T3 + 2) / 4
 h = (T2 + 2*T3 + T4 + 2) / 4
 i = (T3 + 2*T4 + T5 + 2) / 4
 j = (T4 + 2*T5 + T6 + 2) / 4

For RV40 two coefficients differ:

 a = (2*T0 + 2*T1 + L1 + 2*L2 +   L3 +      4) / 8
 f = (  T0 + 2*T1 + T2 +   L2 + 2*L3 + L4 + 4) / 8

For VP8, 3 coefficients differ:

 c = (T2 + T3 + T4 + 2) / 4
 e = (T4 + 2*T5 + T6 + 2) / 4
 j = (T5 + 2*T6 + T7 + 2) / 4

Horizontal/Up

  • H.264: mode 8
  • SVQ3: mode 8
  • RV40: not used
  • VP8: mode 9
    | 
---------------------
 L0 |  a   b   c   d
 L1 |  c   d   e   f
 L2 |  e   f   g   g
 L3 |  g   g   g   g

where:

 a = (L0 + L1 + 1) / 2
 b = (L0 + 2*L1 + L2 + 2) / 4
 c = (L1 + L2 + 1) / 2
 d = (L1 + 2*L2 + L3 + 2) / 4
 e = (L2 + L3 + 1) / 2
 f = (L2 + 3*L3      + 2) / 4
 g = L3

Horizontal/Up (RV40)

  • H.264: not used
  • SVQ3: not used
  • RV40: mode 7
  • VP8: not used
    | T0  T1  T2  T3  T4  T5  T6  T7
-------------------------------------
 L0 |  a   b   c   d
 L1 |  c   d   e   f
 L2 |  e   f   g   h
 L3 |  g   h   i   j
 L4 |
 L5 |
 L6 |

where:

 a = (T1 + 2*T2 + T3 + 2*L0 + 2*L1 +      4) / 8
 b = (T2 + 2*T3 + T4 +   L0 + 2*L1 + L2 + 4) / 8
 c = (T3 + 2*T4 + T5 + 2*L1 + 2*L2 +      4) / 8
 d = (T4 + 2*T5 + T6 +   L1 + 2*L2 + L3 + 4) / 8
 e = (T5 + 2*T6 + T7 + 2*L2 + 2*L3 +      4) / 8
 f = (T6 + 3*T7 +        L2 + 3*L3 +      4) / 8
 g = (T6 +   T7 +        L3 +   L4      + 2) / 4
 h = (                   L3 + 2*L4 + L5 + 2) / 4
 i = (                   L4 +   L5      + 1) / 2
 j = (                   L4 + 2*L5 + L6 + 2) / 4

TrueMotion (VP8)

  • H.264: not used
  • SVQ3: not used
  • RV40: not used
  • VP8: mode 1
 LT | T0  T1  T2  T3
---------------------
 L0 |  a   .   .   .
 L1 |  .   b   .   .
 L2 |  .   .   c   .
 L3 |  .   .   .   d

where this pattern is satisfied:

 a = SATURATE_U8(T0 - LT + L0)
 b = SATURATE_U8(T1 - LT + L1)
 c = SATURATE_U8(T2 - LT + L2)
 d = SATURATE_U8(T3 - LT + L3)

I.e., for each of the 16 samples: (top predictor for column) - (left/top predictor) + (left predictor for row), then saturate in an unsigned byte range 0..255.


16x16 Prediction Modes

DC

  • H.264: mode 0
  • SVQ3: mode 0
  • RV40: mode 0
  • VP8: mode 0

Using the 16 top predictors (T0..T15) and the 16 left predictors (L0..L15), set all 256 elements to the mean, computed as:

if top and left predictors are available
  mean = (sum(T0..T15) + sum(L0..L15) + 16) / 32
else if top predictors are available
  mean = (sum(T0..T15) + 8) / 16
else if left predictors are available
  mean = (sum(L0..L15) + 8) / 16
else
  mean = 128

Vertical

  • H.264: mode 1
  • SVQ3: mode 1
  • RV40: mode 1
  • VP8: mode 1
     | T0  T1  T2  T3  T4  ..  T15
-------------------------- .. -----
     | T0  T1  T2  T3  T4  ..  T15
     | T0  T1  T2  T3  T4  ..  T15
     | T0  T1  T2  T3  T4  ..  T15
 ......
     | T0  T1  T2  T3  T4  ..  T15

Horizontal

  • H.264: mode 2
  • SVQ3: mode 2
  • RV40: mode 2
  • VP8: mode 2
     |
--------------------------- .. -----
  L0 |  L0  L0  L0  L0  L0  ..   L0
  L1 |  L1  L1  L1  L1  L1  ..   L1
  L2 |  L2  L2  L2  L2  L2  ..   L2
 ......
 L15 | L15 L15 L15 L15 L15  ..  L15

Plane

  • H.264: mode 3
  • SVQ3: mode 3
  • RV40: mode 3
  • VP8: not used

Notice that SVQ3 and RV40 follow a slightly different method here.

Given the top predictors (T0..T15), left predictors (L0..L15) and the left-top corner predictor (LT) arranged as follows:

  LT |    T0        T1        T2     ..    T15
-----------------------------------  ..  --------
  L0 | c[ 0, 0]  c[ 1, 0]  c[ 2, 0]  ..  c[15, 0]
  L1 | c[ 0, 1]  c[ 1, 1]  c[ 2, 1]  ..  c[15, 1]
 ......
 L15 | c[ 0,15]  c[ 1,15]  c[ 2,15]  ..  c[15,15]

Compute H' and V':

H' = 1* (T8 - T6) +
     2* (T9 - T5) +
     3*(T10 - T4) +
     4*(T11 - T3) +
     5*(T12 - T2) +
     6*(T13 - T1) +
     7*(T14 - T0) +
     8*(T15 - LT)
V' = 1* (L8 - L6) +
     2* (L9 - L5) +
     3*(L10 - L4) +
     4*(L11 - L3) +
     5*(L12 - L2) +
     6*(L13 - L1) +
     7*(L14 - L0) +
     8*(L15 - LT)

For H.264, compute H and V as:

 H = (5*H' + 32) / 64
 V = (5*V' + 32) / 64

For SVQ3, compute H and V as:

 V = (5*(H'/4)) / 16
 H = (5*(V'/4)) / 16 
 (notice that V and H are computed from H' and V', respectively)

For RV40, compute H and V as:

 H = (5*(H' >> 2)) >> 4
 V = (5*(V' >> 2)) >> 4 
 (like SVQ3 but without swapping and it's important to use shifts here instead of divisions)

The final process for filling in the 16x16 block is:

 a = 16 * (L15 + T15 + 1) - 7*(V+H)
 for (j = 0..15)
   for (i = 0..15)
     b = a + V * j + H * i
     c[i,j] = SATURATE_U8(b / 32)

The SATURATE_U8() function indicates that the result of the operation should be bounded to an unsigned 8-bit range (0..255).

TrueMotion (VP8)

  • H.264: not used
  • SVQ3: not used
  • RV40: not used
  • VP8: mode 3

Given the top predictors (T0..T15), left predictors (L0..L15) and the left-top corner predictor (LT) arranged as follows:

  LT |    T0        T1        T2     ..    T15
-----------------------------------  ..  --------
  L0 | c[ 0, 0]  c[ 1, 0]  c[ 2, 0]  ..  c[15, 0]
  L1 | c[ 0, 1]  c[ 1, 1]  c[ 2, 1]  ..  c[15, 1]
 ......
 L15 | c[ 0,15]  c[ 1,15]  c[ 2,15]  ..  c[15,15]

Compute c[t,l] as:

 c[t,l] = SATURATE_U8(L[l] + T[t] - LT)

I.e., for each element, sum the left and top predictors for the row and column, respectively, and subtract the left-top predictor. Then, saturate the result between 0 and 255.