View
216
Download
0
Tags:
Embed Size (px)
Citation preview
A Sequence-Based Rate Control Framework for Consistent Quality Real-Time Video
Bo Xie and Wenjun Zeng
CSVT 2006
Outline
Introduction Global Bit-allocation Target Bit Calculation Quantization Parameter Determination Simulation Results
Introduction (1)
Rate control Bit-allocation Bit-allocation achievement (QP determination)
MPEG-2 TM5, MPEG-4 Annex L, and H.263+ TMN8 GOP-based bit-allocation Fixed size GOP Constant bit-allocation among GOPs and among the sam
e type frames Some assumptions are necessary
Stationary video sequence Similar characteristics of all GOPs
Fixed-size GOPs do not match the scene structure
Introduction
QP determination Different R-Q models are developed There is no guarantee that a model is
always accurate (model mismatch) Resulting in buffer overflow or underflow
Introduction (3)
Problem Formulation
Subject to
seq
seq1seq
* |)()(|1
minargN
iii
Qqqd-qd
Nq
N
seq,...,2,1 ,),(0 NiBqiB s
TNqqqq },...,,{ **
2*1
*
seq
# of framesDistortion for the ith frame with QP qi
Average distortion
seq
1seq
)(1
)(N
iii qd
Nqd
)0,)(),1(max(),( CqrqiBqiB ii
Actual bits for the ith frame with QP qi Bit rate/Frame rate
Buffer sizeBuffer fullness
introduction (4)
Goal: Constant quality Allocating more bits to high complexity
scenes/frames, and less bits to low complexity scenes/frames Coding complexity: The number of bits required to
encode a frame
Three stages Global bit-allocation model Target bit calculation QP determination
Global Bit-allocation (1)
Relation between rate and complexity
R-J relation
N
ijijS
1
2,
The variance of the ith MB in the jth frameOptimal target bit rate for the jth frame
Bit budget for that GOP
# of pixels for each MB # of MBs for each frame
Average rate for the mth frame (bit/pixel)
The energy of the jth frame
N
iimdev
Nmdev
1
1
A
jiiji xx
Amdev
1, ||
1
Intensity of the jth residue pixel of the ith MB
Average residue intensity of the ith MB
Mean deviation
Average bits for motion vectors of a frame
Global Bit-Allocation (2)
Problems of traditional models They cannot differentiate Intra and Inter
frames (Too less bits for Intra frames) Proposed R-MAD model
Value of original pels for intra MBs
Value of residue pels for inter MBsConstant Shift factor
Is chosen as ½ R = K*(MAD)½
If a frame has MAD > average MAD, and calculated QP < average QP, set QP = average QP
Target Bit Calculation (1)
Calculation of K
represents the complexity of a scene/frame
Scene change detection Stationary assumption is no more necessary Most existing GOP-based bit-allocation schemes use
only past source data Worst case: the scenes get more and more complex or
simpler
Buffer output rate (bit rate/frame rate)
MAD of the current frame
Average MAD of all previous encoded frames
1MAD
n
CK
1MAD
MAD
n
n
Target bit count of the current frame
Target Bit Calculation (2)
Adjustment by actual bit account
Achievement of constant quality Buffer constraint
Target bit count for the ith frameActual bit count for the ith frame
Bs VBV_fullness
Quantization Parameter Determination (1)
Bit-allocation guarantee A R-Q model is used to determine QP for the tar
get bits A traditional model-based QP determination has
no bit-allocation guarantee Resulting in “error propagation” e.g. MPEG-4 Annex L
QP re-adjustment If |Actualbit - Targetbit|/Targetbit > Threshold, re-quant
ization is performed to achieve bit-allocation guarantee
Quantization Parameter Determination (2)
Proposed R-Q model Initial QP determination
(similar to )
2Q
X
MAD
R …
n
Window size
n-M
# of re-quantization timesActual # of bits used for the jth re-quantization of the p*th frame in the past
Actual QP used for the q*th re-quantization of the p*th frame in the past
221
Q
X
Q
X
MAD
R
R-Q model
Most similar MAD
Most similar QP
Quantization Parameter Determination (3)
Proposed R-Q model QP re-adjustment
Check convergence
If re-adjusted QP cannot converge to the target bit count,
MAD/Targetbit
MAD/prevActualbit_prev_new_ QPQP
QP_left QP_rightQP_prev
QP_left QP_right
QP_rightQP_left
Actualbit < Targetbit:
Actualbit > Targetbit:
QP_left QP_right QP_prev
QP_new
Quantization Parameter Determination (4)
Proposed R-Q model Re-quantization algorithm
QP final sanity check If then )QPQP && MAD(MAD 11 nnnn 1QPQP nn
)Targetbit
prevActualbit_prev_new_( QPQP
Quantization Parameter Determination (5)
Proposed R-Q model Buffer overflow checking
Simulation Results (1)
Sequences (QCIF) Foreman (medium motion, one scene
change) Sea World (high motion, two scene
changes) Glasgow (lots of scene cuts) Charlie’s Angels (high motion, lots of
scene changes)
Simulation Results (2)
Model failure rate and re-quantization times
|Actualbit-Targetbit| / Targetbit > 30% MPEG-4 Annex L
Simulation Results (3)
Frame dropping and PSNRMPEG-4 Annex LProposedProposed without re-quantizationR- model*
*Z. He and S. K. Mitra, “A unified rate-distortion analysis framework for transform coding,” CSVT 2001.
Simulation Results (4)
Quality smooth Smaller buffer (0.5s)
Simulation Results (5)
Buffer fullness (0.5s)
Simulation Results (6)
Quality smooth Larger buffer (2s)