CABLE: Cloud-Assisted Bandwidth-efficient LMM-based Encoding for V2X Systems
Researchers have developed CABLE, a novel framework designed to enhance the efficiency of large multimodal models (LMMs) in vehicle-to-everything (V2X) systems. This system reduces communication overhead and cloud-side latency by uploading only masked regions of interest (ROIs) from edge devices to the cloud, rather than full-resolution frames. CABLE utilizes previous segmentation masks, ego-motion compensation, and residual-motion cues to define these ROIs, creating a feedback loop between the cloud and edge. Experiments on multiple datasets demonstrate significant communication savings and improved LMM prefill speeds with only a minor trade-off in detection quality. AI
IMPACT This framework could significantly reduce the computational and communication costs for real-time AI perception in autonomous systems.