EVENT DOSSIER
A study on 360-degree video end-to-end neural compression evaluated the impact of seven projection formats supported by JVET 360Lib on encoding performance. The study utilized a scale-space flow model, JVET test sequences, and conventional conditions to convert the source isometric cylindrical projection into encoding projections for multi-rate point compression and reconstruction, and assessed the results using PSNR, spherical PSNR, weighted spherical PSNR, and Bjontegaard delta rate. The results showed that under the scale-space flow model, the isometric and filled-isometric cylindrical projections provided the highest compression efficiency; in contrast, cube map and菱形 dodecahedron projections performed poorly, which differed from the better performance of cube maps in traditional HM-16.16 encoders. Additionally, the optical-flow-based neural model performed well due to the spatial continuity of single-sided projections, while the block-based hybrid encoder was more suitable for multi-sided layouts. The study conclusions indicate that projection efficiency depends on the type of encoder, providing guidance for selecting appropriate projections for 360-degree video compression.
本研究评估了七种由 JVET 360Lib 支持的投影格式对 360 度视频端到端神经压缩的影响。研究利用尺度空间流模型、JVET 测试序列及常规测试条件，将源等距圆柱投影转换为编码投影进行多速率点压缩与重构，并通过 PSNR、球面 PSNR、加权球面 PSNR 及 Bj{\o}ntegaard delta 率评估性能。结果显示，在尺度空间流模型下，等距圆柱和填充等距圆柱投影提供最高压缩效率，而立方体贴图和菱形十二面体投影效果较差；这与传统 HM-16.16 编码器中立方体贴图格式表现更优的情况不同。此外，基于光流的神经模型受益于单面投影的空间连续性，而基于分块的混合编码器则更适合多面布局。研究结论表明投影效率依赖于编码器类型，为学习基 360 度视频压缩的投影选择提供了指导。