Event Camera Monocular Depth Estimation: An Event-Tailored State-Space Model-Based Approach

Event cameras, characterized by their high temporal resolution, wide dynamic range, and inherent immunity to motion blur, are emerging as a powerful and robust sensing modality for depth measurement in dynamic and challenging environments. While existing methods have made progress by adapting image-domain networks like RNNs or Transformers, they generally do not exploit the unique asynchronous, sparse, and temporal nature of event streams. To this end, we propose ESDepth, a pure state-space model specifically architected for monocular depth estimation with event cameras. The key innovation of ESDepth is the Event State-Space Model (ESSM) architecture, which incorporates modules tailored for event data characteristics, including a Multi-Frequency State-Space Block (MFSSB) for efficient time-frequency encoding of event data and an Event Distribution Gating Unit (EDGU) with event priors for effective modulation through gating mechanisms. After achieving time-frequency encoding, the vision-based selective state-space model is applied to encode the spatial features of event data, which finally enters a decoder to produce the desired dense depth map. In summary, guided by event priors, the ESDepth framework comprehensively incorporates temporal, frequency, and spatial information within the state-space model, enhancing its ability to adapt to asynchronous and sparse event data. Designed with measurement accuracy and efficiency in mind, ESDepth achieves the best overall performance across multiple driving scenarios, demonstrating its potential for real-time depth sensing and dynamic measurement applications. The code is available at: https://github.com/Kingtorah/ESDepth.