Memory-Aware ROI-Based Power-Efficient 360-Degree Video Processing

This paper presents a memory-aware optimization framework for power-efficient 360-degree video processing. A Python-based pipeline is developed to analyze region-of-interest (ROI)-guided data optimization with YOLO model, aligned with human visual perception. While conventional encoders compress video uniformly, immersive 360-degree content contains regions with varying perceptual importance. We propose an adaptive precision approach in which ROI regions are processed at full precision, while non-ROI regions undergo controlled bit truncation to reduce memory usage and power consumption. To support this strategy, a flexible heterogeneous memory partitioning scheme is introduced to accommodate dynamic ROI-to-non-ROI ratios. A 6T SRAM-based architecture is designed and verified in 65 nm CMOS technology from TSMC at 1.2 V. Cadence-based simulation results demonstrate approximately 22% memory-related (read) power savings by 3-bit truncation per byte with negligible perceptual degradation, which enables efficient immersive video processing for resource-constrained edge systems. Additionally, on average 35% bits are saved per frame, and the PNG size is optimized by 51% on Timelapse video. The cumulative power number is significant for complete video processing. Moreover, our results show that on average hardware-based per-pixel bit truncation is four orders (1000 x) of magnitude faster than software-based truncation approach.