TL;DR
This paper tackles the complexity of Spatial Scalable High Efficiency Video Coding (SSHVC) by proposing a novel, faster intra-prediction method. By observing that the rate-distortion costs of Inter-layer Reference (ILR) and Intra modes follow Gaussian distributions, they introduce a Bayes decision rule to efficiently skip unnecessary Intra mode evaluations, significantly speeding up encoding with minimal loss in quality.
What Problem Does This Paper Address?
The increasing demand for video streaming across diverse devices with varying resolutions has led to the development of Spatial Scalable High Efficiency Video Coding (SSHVC). SSHVC allows for adaptive decoding of different layers to match display capabilities. However, a major drawback of SSHVC is its inherent complexity, particularly in the encoding process. This complexity arises from the need to evaluate additional Inter-layer Reference (ILR) modes, which predict a coding unit (CU) in an enhancement layer (EL) from a corresponding CU in the base layer (BL). This complexity hinders the widespread adoption of SSHVC, especially in real-time and wireless applications. The paper aims to address this by developing a faster intra-prediction algorithm that reduces encoding time without sacrificing coding efficiency.
Methods
The core of the proposed method lies in a novel mode selection strategy. The authors first investigate the rate-distortion (RD) costs of ILR mode and Intra mode for different coding depth levels in SSHVC. They observe that these RD costs, while distinct, consistently follow a Gaussian distribution. This observation is crucial.
Leveraging this Gaussian distribution property, they propose a Bayes decision rule. The rule aims to predict whether the ILR mode is the optimal choice for a given CU. If the probability of ILR mode being the best is sufficiently high (e.g., above 0.95), the algorithm directly skips the evaluation of Intra modes for that CU. This early termination significantly reduces the computational burden. The thresholds for this decision (To and T1) are derived based on the conditional probability density functions of the RD costs for both ILR and Intra modes, utilizing the properties of the Gaussian distribution.
Datasets
The experiments were conducted using several video sequences: Blue_sky, Ducks, Pedestrian, Tractor, Town, and Station2. These sequences were chosen for their diverse motion and texture characteristics, ranging from simple to complex. The evaluation was performed using common SHM test conditions (CSTC) with specific QP (Quantization Parameter) settings for the base layer (BL) and enhancement layers (ELs).
Key Insights and Ideas
- Gaussian Distribution of RD Costs: The most significant insight is the consistent observation that the rate-distortion costs for both ILR and Intra prediction modes in SSHVC follow a Gaussian distribution across various depth levels and video sequences. This statistical property is the foundation for the proposed optimization.
- ILR vs. Intra Mode Performance: The paper highlights that the performance of ILR and Intra modes is context-dependent. ILR mode often outperforms Intra mode for CUs with complex textures at coarser depth levels (0-2) due to its ability to leverage the base layer's information. Conversely, for simpler textures or at finer depth levels (3), Intra mode can be more efficient due to local spatial correlations.
- Bayes Decision Rule for Early Skipping: The application of a Bayes decision rule, informed by the Gaussian distribution of RD costs, provides an effective mechanism for early skipping of Intra mode evaluations. This is a direct consequence of the statistical insight.
- Trade-off between Speed and Efficiency: The research demonstrates that it's possible to achieve substantial speed improvements in SSHVC encoding without a significant degradation in coding efficiency, challenging the notion that complexity is an unavoidable trade-off.
Key Results
The proposed Gaussian Distribution-based Mode Selection algorithm demonstrated significant improvements in coding speed.
- Coding Speed (TS): The average encoding speedup (TS) for the enhancement layer (EL) was reported as 55.85% in case 1 and 58.91% in case 2. This is considerably higher than the compared methods, FIICA (42.22% and 43.25%) and PBFIP (41.27% and 47.81%).
- Coding Efficiency (BDBR): The algorithm achieved very low Bjontegaard Delta Bit Rate (BDBR) values, indicating negligible coding efficiency losses. The average BDBRs were 0.26% (case 1) and 0.24% (case 2), which were lower than or comparable to the other methods.
Limitations and Future Work
- Limitations: The paper focuses specifically on intra-prediction for SSHVC. While the results are promising, the direct applicability to other video coding standards or different prediction types (e.g., inter-prediction) is not explicitly discussed. The effectiveness of the Gaussian assumption might also be sensitive to extreme outliers or highly non-standard video content.
- Future Work: The authors suggest that further investigation into the fundamentals of mode selection could lead to more advanced algorithms. Exploring the application of this Bayes decision rule to other aspects of video coding or different scalable extensions could be a natural next step. Additionally, investigating the robustness of the Gaussian assumption under various conditions and for different sequences would be valuable.
Summary updated