Test for Error Recovery Capability

To verify the error recovery capability of the RFGS, a simple experiment is performed to demonstrate the worst-case scenario when there is bandwidth variation that can result in maximal effect of drift. We assume the network bandwidth is sharply dropped for every first P-picture transmitted of each GOV and the bit budget for the other frames is set as 1024 kbps. Such a bandwidth scenario is illustrated in Figure 3.14.

Since only the first P-picture for the enhancement layer is lost and the degradation of the subsequent frames will be caused only by the errors from this P-picture. The same testing conditions and the video sequences are used as in [17]. To verify the error attenuation of RFGS mentioned in the Section 3.3.3, we first examine the RFGS1 method about the speed of the error recovery for various α. In all the simulations, β is set as 3 and α equals to one of the four predefined values, 0.5, 0.75, 0.9, and 1.0. As shown in Figure 3.15, the error attenuation capability of the RFGS framework is strongly affected by the value of α used. At the worse case scenario that no enhancement bit is received, the PSNR loss is more than 5 dB as compared to the PSNR under an error-free condition. For a small α of 0.5, the error is attenuated very fast. For example, in Figure 3.15, after fourth P-pictures within the first GOV, the PSNR differences are reduced to about 0.1 to 0.3 dB. When α equals to unity, as shown in the fourth GOV in Figure 3.15, the drift lasts for a long time. We provide the performance

-4 -2 0 2 4 6 8 10

0 50 100 150 200 250 300

Frame Index

PSNR (dB)

RFGS(0.50, 3) RFGS(0.75, 3) RFGS(0.90, 3) RFGS(1.00, 3)

Figure 3.15.The error attenuation in PSNR for the Y component of the Akiyo sequence under different α in the RFGS1 framework, where the pair of the values indicates the prediction mode parameters

(

)

of RFGS2_LM under the burst error in Figure 3.16. We simulate the burst error with a loss of the first few frames in every GOV. Two burst lengths of one frame and seven frames are used for simulation. By applying the RFGS method for both the enhancement and base layers, the error drift is more serious as compared the drift for the RFGS1. However, the visual quality can still be fast recovered from the burst errors.

We also perform the dynamic test following the channel bandwidth variation pattern as defined in [19] to demonstrate the performance of RFGS. The bandwidth pattern as illustrated in Figure 3.2 are as follows. The total bandwidth is switched in a step size of 256 kbps that decreases from 1024 kbps to 256kbps and increases back to 1024 kbps. The instantaneous bitrate is held for 24 seconds (or 720 frames with frame rate 30). Other test conditions are identical to those described in Section 3.5.1 and as defined in [17]. In the simulation, the Novel sequence in CIF format and with the frame rate of 30 fps were used. The first 5040 frames of the sequence are used for testing and the base layer is coded at 256 kbps. During transmission, we use the 2-Level Priority

25 27 29 31 33 35 37 39 41 43

0 60 120 180 240 300

Frame Index

PSNR (dB)

RFGS2_LM None Drop RFGS2_LM Drop 1 RFGS2_LM Drop 7 BaseLineFGS_drop=7

Figure 3.16. The error attenuation in PSNR for the Y component of the Foreman sequence using the RFGS2_LM framework. All the curves denote truncation of the enhancement layer bitstream at 1024kbps. For the curve labeled ‘RFGS2_LM Drop 1’, the first frame of each GOV is dropped. For the curve labeled ‘RFGS2_LM Drop 7’, the first seven frames of each GOV are dropped. For the curve labeled ‘RFGS2_LM None Drop’, no frame is dropped. The curve labeled ‘BaseLineFGS_drop=7’ is the baseline FGS with the first 7 frames of each GOV dropped.

Network, where the FGS base-layer is set at high priority. When the bandwidth is small, the base layer will be sent first. For the single layer approach, we encode the bitstream with 256kbps, 512kbps, 768kbps, and 1024kbps and dynamically select the appropriate bitstreams for the target bitrates as defined in [19].

Figure 3.18 shows the simulation results. As compared with the results based on the

single layer and the baseline FGS approaches, the results show that the RFGS2 with the linear model can adaptively select the suitable α offline to achieve similar performance as that of the single layer approach for given dynamic bandwidths and different scene over a long sequence.

As for the error recovery speed for different sequences, as shown in Figure 3.17, it is observed that the error recovery is also related to the temporal dependency between the successive frames of the same sequence. For the fast moving sequences like Coastguard and Foreman, the current frame only refers to a fraction of information from

0 5 10 15

0.4 0.6 0.8 1 Alpha value

Time constant

Akiyo Coastguard Foreman

Figure 3.17.The relationship between the leak factor α and the time constant τ for the error attenuation. For each curve, β is 3.

the reference frame, which allows limited error propagation. Thus, the errors vanish even with a larger leak factor α. For the slow motion sequences such as Akiyo, most of the frames consist of static areas such that there exist strong dependencies between the consecutive frames of the sequence. The dependencies can improve the coding efficiency but suffers from more drift when the transmission bandwidth is insufficient.

Therefore, the RFGS with a small α (about 0.5) is recommended for the slow motion video sequences to improve the error robustness.