Digital Signal Processing Reference
In-Depth Information
Table 1.
Main simulation parameter settings
Parameters Values
Target bitrate 128 kbps(
Akiyo
), 256 kbps (
Foreman
), 384 kbps (
Football
)
Target frame rate 15 fps
Codec H.264 JM17.0, Main profile, Level 2
GOP structure IPPP···IPPP···
GOP length (
N
g
) 15 frames
Intra refresh Off
Rate control Enable
Slice mode Fixed macroblock number in a slice
η
1, 2, 3, 4
Streaming protocols RTP/UDP/IP
Channel model Two-state Markov
p
h
0%
p
b
0%, 1%, 3%, 5%, 10%, 15%, 20%
Here we examine and compare the loss-resilience performance of three online
packet classification schemes: (a) the traditional in-order packet classification (IPC).
In encoding order, the former
η
packets of each frame are classified as the high-
reliability traffic class; (b) the content-aware online packet classification (COPC). The
COPC scheme is similar to the classification mechanism in [6]; (c) The proposed
propagation-aware online packet classification (POPC).
Fig. 3 depicts the PSNR(
p
b
) curve as a function of the best-effort packet loss
probability
p
b
. We can see that the POPC scheme consistently outperforms two
reference schemes for different
p
b
values. Since the IPC scheme only takes into
account the spatial-temporal position of a packet in a GOP, some important packets
cannot avoid the losses which result in significant PSNR(
p
b
) drop. For very low
p
b
=1%, the loss-resilience performance is very close due to the isolated error effect.
For
p
b
=3%, the PSNR(
p
b
) differences of the POPC and COPC schemes are relatively
insignificant. For
p
b
=5-15%, the PSNR(
p
b
) differences of the POPC and COPC
schemes are relatively significant. We can also see that PSNR(
p
b
) drops down sharply
in the range of
p
b
=15-20%. The reason is that burst packet losses in this
p
b
range often
result in the whole-frame losses, and the error concealment in the JM decoder cannot
recovery whole-frame loss well.
For three online packet classification schemes, Fig. 4 further demonstrates an
empirical cumulative distribution function (CDF) of the frame-by-frame PSNR in all
simulations. Since the issue of time-varying error propagation between frames is still
ignored, the distortion estimation of the COPC scheme cannot adapt the actual
conditions well, especially for serious packet losses and high-dynamic video content.
Since the GOP-level transmission distortion is effectively estimated, the POPC
scheme is more accurate than other two online classification schemes, and thus
achieves better loss-resilience performance.