Enhancing Feature Coding Efficiency via Global Statistics Preservation

This presentation delves into the emergence of AI-enabled edge devices and the split-inference paradigm, emphasizing the importance of feature compression. We discuss the Feature Coding for Machines (FCM) pipeline, challenges it faces, and innovations under MPEG's standards. A proposed method preserving global statistics via Z-score normalization and periodic signaling is outlined. Experimental results highlight improvements in instance segmentation, bitrate reduction, and object tracking....