Neural Latitude-Aware Adaptive Coding for Efficient 360° Video Representation

Omnidirectional (360°) videos have paved way for increasingly immersive experiences in applications including panoramic streaming as well as virtual reality. But the conventional coding pipelines for videos uniformly treat all the spatial regions. This happens despite the highly non-uniform sampling and perceptual importance produced by spherical projections such as equirectangular mapping. This paper attempts to introduce a neural latitude-aware adaptive coding framework which tries to dynamically adjust the coding behavior based on the spatial latitude information. Creating a lightweight neural model that predicts optimal per-latitude-band (32-band) adaptation factors, the proposed method allocates higher fidelity to perceptually critical regions and reduces redundancy in regions found to be less significant. The approach is evaluated on a diverse set of 360 video sequences and is compared against a fixed latitude-based baseline.