TL;DR
This paper introduces a novel 3D video compression algorithm that leverages wavelet transforms and the EBCOT (Embedded Block Coding with Optimal Truncation) algorithm to achieve both SNR and resolution scalability. The proposed method offers comparable compression efficiency to MPEG-4 while providing a more linear and flexible scalable bitstream, outperforming the 3-D SPIHT algorithm by 1.5dB.
What Problem Does This Paper Address?
Transmitting video over packet-switched networks presents a significant challenge due to varying network conditions. Traditional video compression methods often lack the flexibility to adapt to these changes, leading to inefficient data usage or poor quality. Scalable video compression aims to address this by allowing a single bitstream to be decoded at different quality levels, frame rates, or resolutions. While existing methods like MPEG-4 offer some scalability, they can suffer from a "staircase" improvement in quality and potential drifting issues in predictive coding. This paper seeks to develop a more robust and flexible scalable video compression scheme that overcomes these limitations.
Methods
The core of the proposed algorithm lies in a two-stage wavelet decomposition followed by EBCOT compression.
- 3D Wavelet Decomposition: The video sequence is first decomposed spatially using biorthogonal Antonini 9/7 filters. This is then followed by a temporal decomposition using the Haar wavelet. This 3D spatio-temporal decomposition is applied to groups of frames (GOFs), with the authors experimenting with GOF sizes ranging from 4 to 32. The use of wavelets helps avoid the blocking artifacts common in DCT-based methods.
- EBCOT Compression: The resulting wavelet coefficients are then compressed using the EBCOT algorithm, which is a key component of the JPEG2000 standard. EBCOT is known for its ability to generate highly scalable bitstreams by partitioning subbands into code blocks and creating an embedded bitstream for each. This allows for fine-grained control over the bitstream truncation, enabling flexible rate scalability. The authors utilize the Kakadu software for EBCOT implementation.
Datasets
The paper evaluates the proposed algorithm using the "Foreman" video sequence. No new datasets were introduced.
Key Insights and Ideas
- Wavelets for Scalability: The inherent multi-resolution nature of wavelet transforms is a strong foundation for achieving both spatial and temporal scalability in video compression.
- EBCOT's Granular Scalability: EBCOT's ability to generate an embedded bitstream for each code block is crucial for achieving fine-grained rate scalability, allowing for a smooth increase in quality rather than discrete steps.
- 3D Decomposition Benefits: Applying wavelet decomposition in both spatial and temporal domains effectively captures redundancies and enables efficient compression while maintaining scalability.
- Resilience to Errors: The paper notes that wavelet-based approaches, due to their block-based nature (even with wavelets), can offer better resilience to bit errors compared to traditional hybrid coding schemes where errors can propagate. However, error resilience is not the focus of this particular study.
Key Results
- Performance vs. MPEG-4: The proposed algorithm achieves compression performance comparable to MPEG-4. Figure 4 shows that the proposed scheme performs better in most frames of the Foreman sequence at a bit rate of 0.15 bpp.
- Performance vs. 3D SPIHT: The algorithm demonstrates superior performance compared to the 3-D SPIHT algorithm, achieving approximately 1.5dB better PSNR (Figure 5).
- Scalability: The EBCOT-based algorithm provides a more "linear scalable output stream" compared to MPEG-4's enhancement layer allocation, offering up to 30 possible quality layers.
- Computational Complexity: The execution time for the EBCOT video coding is reported to be almost constant at around 20 seconds (on a 600MHz CPU), which is significantly more stable than MPEG-4's execution time, which varies with compression ratios. This constant time is beneficial for real-time rate adaptability.
- GOF Impact: The study found that increasing the GOF size beyond a certain point (around 16 frames) led to a decrease in average PSNR, suggesting a trade-off between temporal correlation capture and sensitivity to motion changes.
Limitations and Future Work
- Error Effects Not Considered: The paper explicitly states that the effects of bit errors on the compressed stream are not considered in this work.
- Computational Cost of 3D Wavelets: While EBCOT's encoding/decoding time is stable, the initial 3D wavelet decomposition itself can introduce delays and require larger buffer sizes, which is a known characteristic of 3D wavelet approaches.
- Future Work: The authors suggest that the constant encoding and decoding time makes this algorithm a strong candidate for real-time rate adaptation over packet-switched networks. Further exploration of error resilience could be a valuable future direction.
Summary updated