Sensor-Algorithm Co-Design for Spectral Sampling in Neuromorphic Vision

Neuromorphic event cameras encode visual information as sparse asynchronous signals resembling spike-based sensory coding in biological systems. While most event-based perception methods rely on grayscale signals and supervised learning, the computational role of spectral diversity in neuromorphic sensing remains largely unexplored. We propose a task-driven formulation in which spectral sampling structure is treated as an optimizable component of the sensing pipeline. Color event streams generated by a Bayer-pattern sensor are represented as spatially interleaved multi-channel spike signals and processed using a compact unsupervised discriminative correlation filter network trained via forward-backward consistency. A learnable mosaic layer enables joint optimization of spectral sampling and downstream task consistency. Experiments across multiple temporal resolutions show that improvements arise primarily from enhanced spectral representation rather than increased model capacity. These results support a sensor-algorithm co-design perspective in which adaptive sampling structures improve information efficiency in event-based perception.