Sadiq (Layi) Macaulay

Learning depth from event-camera streams remains challenging because events are irregular in time, while supervision is usually available only as discretized depth maps and camera poses at sparse reference timestamps. Existing methods sidestep this by collapsing events into synchronous frames, leaving continuous per-event depth unexplored. We introduce the first framework to predict depth per event as each arrives with low latency. It builds on deep state-space models that learn latent dynamics from irregularly sampled observations.

To enable learning under sparse supervision, we formulate a temporally aligned geometric supervision strategy based on differentiable event warping. Experiments on event-based depth benchmarks demonstrate that the proposed framework improves accuracy and temporal consistency over prior depth methods.

Event cameras measure brightness changes asynchronously, providing microsecond timing, high dynamic range, and low-latency visual sensing. These properties make them attractive for depth estimation in fast and low-light scenes, but dense depth labels are not naturally defined at event times and are not required by our training objective. Many self-supervised pipelines nevertheless collapse the stream into frame-like representations, reducing the fine-grained timing that makes event sensing valuable. We present EvSSE, an event-native continuous-time framework for self-supervised per-event depth prediction from asynchronous events. A selective state-space encoder with local spatiotemporal aggregation predicts per-event depth. Training is driven by motion-compensated event alignment and regularization rather than dense depth annotation. Dense depth at any target time is rendered by continuous-time warping and the proposed occlusion-aware TODepth resampler. Experiments show that EvSSE preserves event timing through the depth pipeline while improving event-only depth estimation

Event cameras, with their high dynamic range (HDR) and low latency, offer a promising alternative for robust depth estimation in challenging environments. However, many event-based depth estimation approaches are constrained by small-scale annotated datasets, limiting their generalizability to real-world scenarios. To bridge this gap, we introduce EvtSlowTV, a large-scale event camera dataset curated from publicly available YouTube footage, which contains more than 13B events across various environmental conditions and motions, including seasonal hiking, flying, scenic driving, and underwater exploration. EvtSlowTV is an order of magnitude larger than existing event datasets, providing an unconstrained, naturalistic setting for event-based depth learning. This work shows the suitability of EvtSlowTV for a self-supervised learning framework to capitalise on the HDR potential of raw event streams. We further demonstrate that training with EvtSlowTV enhances the model's ability to generalise to complex scenes and motions. Our approach removes the need for frame-based annotations and preserves the asynchronous nature of event data.