EMDepth: Self-Supervised Event-Based Monocular Depth Estimation

Event camera, a novel bionic vision sensor, fundamentally diverge from traditional imaging devices in their operational principles. Featuring high temporal resolution and extensive dynamic range, these sensors demonstrate exceptional performance in challenging imaging scenarios. However, the event streams generated by such cameras are inherently asynchronous, sparse, low spatial resolution and signal-to-noise ratios, rendering direct application in conventional depth estimation methodologies problematic. In this study, we introduce EMDepth, a innovative monocular event camera-based scene depth estimation method. EMDepth innovatively preserves the spatial and temporal neighborhood characteristics of event points, enabling accurate scene depth reconstruction under low-light conditions and rapid motion scenarios. Experimental validation on the DSEC dataset substantiates the method's effectiveness, demonstrating superior performance compared to current supervised event-driven and unsupervised image-based approaches.