HiFiC360: High-Fidelity Compression for 360° Images

Abstract

360° images, defined on a sphere, are widely used in immersive applications, but their equirectangular projection introduces distortions, especially near the poles, while the equatorial region is most perceptually important. This paper adapts the HiFiC model to ${360}^{{\circ}}$ imagery through three changes: (i) replacing LPIPS with a weighted-to-spherically-uniform similarity loss, (ii) considering spherically-weighted horizontal dilated convolutions, and (iii) adjusting cropping (used to reduce memory usage) to emphasize informative regions. Trained on SUN360 and evaluated on standard benchmarks, HiFiC360 significantly improves the baseline (achieving up to −24.32% BD-Rate) at similar bitrates and training cost, outperforming prior neural methods in weightedto-spherically-uniform peak signal-to-noise ratio across bitrates.