Abstract
This paper introduces a novel interactive compression scheme for 360-degree images designed to achieve high transmission efficiency without the prohibitive storage overhead of pre-encoding every possible viewport. By utilizing rate-adaptive Low-Density Parity-Check (RA-LDPC) codes, the system allows for fine-grained data extraction tailored to a user's current viewport. This approach maintains high compression by exploiting visual correlations between image blocks regardless of their access order, effectively bridging the gap between standard tiling methods and theoretical "oracle" transmission rates.
Introduction
Omnidirectional (360-degree) images are essential for immersive virtual reality experiences but present significant bandwidth challenges, often requiring 8K resolutions to maintain visual quality. When viewing these images through a Head-Mounted Display (HMD), a user only sees a small portion of the sphere at any given time, known as the viewport.
Traditional compression methods face a fundamental trade-off: high compression efficiency relies on predictive coding, which creates complex data dependencies, while random access for bandwidth saving usually requires "tiling" (partitioning the image into independent blocks). Tiling often hurts compression by ignoring cross-tile correlations and results in wasted bandwidth when tiles are larger than the requested viewport. This research proposes a "best of both worlds" solution using incremental codes to provide fine-grained access with minimal storage impact.
System Architecture and Methodology
The core of the proposed system is the use of rate-adaptive channel codes (RA-LDPC) for source coding. Unlike standard codecs like HEVC or VVC that enforce a rigid raster-scan decoding order, this framework allows blocks to be decoded in various sequences based on the user's specific request.
As shown in the system architecture, the process begins with mapping the spherical image to a 2D representation (e.g., Equirectangular or CubeMap). The encoder generates bitstreams based on a "worst-case" prediction scenario. When a user requests a specific viewport (defined by $\theta$), the server extracts only the necessary bitstream subsets. Because the decoder can use previously decoded neighboring blocks to form a prediction, the server only needs to send enough data to "correct" that prediction into the original block. This makes the bitstream inherently elastic and highly efficient.
Optimized Scanning and Access
To ensure that any user request has a valid starting point for decoding, the authors developed strategic "access block" placement algorithms. These standalone blocks can be decoded without reference to neighbors. Once an access block is established, the system must traverse the rest of the viewport efficiently.
The researchers introduced a SnakeLike scanning algorithm, illustrated above. This algorithm traverses the requested viewport in a manner that maximizes horizontal predictions. Their analysis indicates that for 360-degree content, horizontal correlations are generally stronger and more efficient to exploit than vertical ones. This fine-grained, block-level approach avoids the "staircase effect" common in tile-based systems, where bit rates spike abruptly as a user's gaze crosses tile boundaries.
Key Results and Performance
The proposed RA-LDPC coder was evaluated using the Nantes University Dataset and real-world user head-movement trajectories. Key findings include:
- Transmission Efficiency: The method achieves bit-rate savings of approximately 84% compared to non-tiled approaches for initial viewport requests, nearly reaching the performance of an "oracle" coder.
- Storage Overhead: While storing every possible viewport exhaustively would increase storage needs by over 1100%, this method only adds a modest 56% overhead to the base image storage.
- Usefulness: The system demonstrates a high "usefulness" metric, meaning the ratio of pixels actually displayed to the user versus those decoded is significantly higher than in traditional tiling methods.
Conclusion and Future Work
This research demonstrates that rate-adaptive channel codes can effectively decouple data storage from consumption order in 360-degree imaging. While the computational complexity of RA-LDPC decoding remains higher than standard arithmetic coding (like CABAC), it offers a viable path toward extremely efficient interactive streaming. Future work will focus on optimizing these algorithms for real-time performance on mobile HMDs and extending the fine-granularity access model to 360-degree video, where temporal dependencies add a further layer of complexity.
Summary updated