New neuromorphic vision method runs real-time binarization on CPUs

A team of researchers (Pei Zhang, Shijie Lin, Zhou Ge, Jinpeng Chen, Wei Pu) has published a paper on arXiv (2605.17984) demonstrating a novel neuromorphic vision technique that can binarize silhouettes in real time using only CPU processing. The method leverages event cameras—which offer microsecond-level temporal resolution and high dynamic range—to overcome the motion blur and lighting challenges that plague traditional frame-based imaging on mobile platforms like drones, autonomous vehicles, and underwater robots. Their dual-modal approach fuses frames and events asynchronously, avoiding the common issue of event scarcity that breaks time-binning reconstruction. This allows the system to maintain clear target shapes even at extreme kilohertz frame rates, while running efficiently on edge devices with no GPU required.

The work focuses on binarization of quasi-bimodal objects (text, road signs, barcodes) essential for visual communication. Extensive evaluations show competitive performance against leading techniques, with impressive improvements under challenging illumination. The binary outputs serve as reliable representations for downstream tasks such as object detection, tracking, and navigation. This paves the way for lightweight perception in embodied intelligence on resource-constrained edge platforms—think drones avoiding obstacles at high speed or self-driving cars reading road signs in a blizzard. The 12-page paper includes 12 figures and 3 tables, and is currently under review, with a project page linked in the preprint.