Forward channel Performance evaluation
14/09/2010 - Results were updated !
The performance evaluation of the forward channel is presented here, this means the channel used by the encoder to transmit to the decoder the bitstream of the key frames (H.264/AVC Intra bits) and the syndrome bits for the WZ frames (DVC parity bits).
Since the forward channel bits are sent to reach the maximum video quality for each selected quantization matrix (see test conditions), this study mainly regards the evaluation of the overall video quality and compression efficiency, this means both for the WZ and key frames bitstreams. The quality metric used is the average PSNR over all the frames of a sequence coded for a certain quantization matrix.
Overall Rate-distortion performance
The RD performance of the DISCOVER codec will be compared with the corresponding performance of three standard coding solutions which share an important property in terms of encoder complexity: the complex and expensive motion estimation task is not performed by any of them. The three state-of-the-art standard solutions used here for comparison purposes are:
- H.263+ Intra: Video coding with H.263+ without exploiting temporal redundancy; this ‘Intra comparison’ is still the one that appears most in the DVC literature but it is clearly not anymore the best standard Intra coding available and thus ‘beating’ H.263+ Intra is much easier than ‘beating’ H.264/AVC Intra (of course, their encoding complexity is also rather different).
- H.264/AVC Intra: Video coding with H.264/AVC in Main profile without exploiting temporal redundancy; this type of Intra coding is the most efficient Intra (video) coding standard solution, even more than JPEG2000 for many conditions. However, it is important to notice that H.264/AVC Intra exploits quite efficiently the spatial correlation (at a higher complexity cost when compared to H.263+ Intra) with several 4×4 and 16×16 Intra modes, a feature that is (still) missing in the DISCOVER codec.
- H.264/AVC Inter No Motion: Video coding with H.264/AVC in Main profile exploiting temporal redundancy in a IB…IB… structure but without performing any motion estimation which is the most computationally expensive encoding task. The so-called “no motion” mode achieves better performance than Intra coding because it can partly exploit temporal redundancy by using a DPCM temporal scheme, but requires far less complexity than full motion compensated Inter coding because no motion search is performed.
The following results were obtained:
1. GOP 2 RD performance results
2. GOP 4 RD performance results
3. GOP 8 RD performance results
4. GOP 2, 4 and 8 RD performance comparison
Measuring the Quality Evolution of WZ Decoded Frames
This PSNR temporal evolution of the decoded luminance data is also evaluated. For a better understanding of the results, it is important to mention that, for each Qi, there are DCT coefficients bands quantized with ’0’ which means that no parity (WZ) bits are sent to improve the quality (or fidelity) of these coefficients bands. In these cases, the decoder just uses the side information bands which still have errors and, therefore, error propagation at the decoder can occur; this is mitigated when some bands/bitplanes are correctly decoded (with a low error probability). The higher is Qi, the less important is this effect.
The following results were obtained:
Comparing the RD Performance of LDPC and Turbo Codes
During most the DISCOVER project, the Slepian-Wolf codec part of the monoview codec was based on turbo codes. Later, turbo codes were substituted by LDPC codes, essentially due to their better RD and complexity performance, especially at the decoder side.
The following results were obtained:
6. RD performance comparison using Turbo Codes and LDPC codes
Measuring the Bitplane Compression Factor
The total number of parity bits per bitplane created by the LDPCA encoder is equal to the number of the input bitplane bits. This way the compression factor can vary between 0 and 1 (but never be lower than 1 as for turbo codes) and allows to measure the capability of the DISCOVER codec to compress the source data (bitplanes). The compression factor is defined in [1].
The following results were obtained:
7. Bitplane Compression Factor
Measuring the Decoded Quality Versus the Side Information Quality
The PSNR gain obtained after WZ decoding with respect to the side information is evaluated here. This gain is defined as the difference between the PSNR of the WZ frames and the PSNR of the corresponding side information. While it is obvious that this gain is always positive, it is important to understand how big these gains are and for what conditions.
The following results were obtained:
8. WZ frames versus Side Information PSNR
References
1. J. Pedro, C. Brites, J. Ascenso, F. Pereira, "Studying the Feedback Channel in Transform Domain Wyner-Ziv Video Coding", 6th Conference on Telecommunications - ConfTele, Peniche, Portugal, 2007.