Researchers have developed Mondrian, a new edge system designed for high-performance object detection in high-resolution video streams. The system employs a novel Compressive Packed Inference technique to reduce per-pixel processing costs by selectively processing necessary pixels and maximizing parallelism. Mondrian dynamically extracts and shrinks Regions of Interest (ROIs), then intelligently combines them to optimize the utilization of graphics processing units (GPUs) and similar accelerators. Evaluations show Mondrian achieves 15.0-19.7% higher accuracy and up to 6.65x higher throughput compared to existing methods for processing 1080p video. AI
IMPACT This system could enable more efficient and powerful on-device video analysis for applications like autonomous systems and surveillance.
RANK_REASON This is a research paper detailing a new system and technique for video analytics. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
