New Graph Spectral Denoising Method Cleans Up Neuromorphic Camera Noise

Neuromorphic cameras detect brightness changes asynchronously per pixel, outputting 3-D streaming data (pixel coordinates + time). While offering high temporal resolution, low latency, and low power consumption, their high sensitivity introduces significant noise. Shimpei Harada, Junya Hara, Hiroshi Higashi, and Yuichi Tanaka from their respective institutions propose a new denoising approach based on graph spectral features. The method builds a graph where each node is an event and edges encode spatiotemporal distance. Using a prior on the density of 3-D events, they calculate a graph-specified parameter that controls connectivity. They then compute eigenvectors of the graph Laplacian to directly extract noiseless events.

A key innovation is the customization of the graph Laplacian to reorder its eigenvalues, enabling fast eigensolver algorithms instead of naive eigendecomposition. This drastically reduces computational complexity. In experiments on both synthetic and real-world event data, the proposed method effectively removes noise events and outperforms alternative denoising techniques. This work, submitted to arXiv on May 14, 2026, addresses a critical bottleneck for neuromorphic vision systems, which are increasingly used in robotics, autonomous driving, and high-speed imaging. By cleaning noisy event streams efficiently, it brings practical deployment one step closer.