Event-Based Optical Flow

Event-based optical flow recovers how image points move over time from cameras that report brightness changes asynchronously rather than capturing conventional frames. Methods encode events as time surfaces or occupancy grids, then use recurrent networks, trajectory-based feature warping, or spiking models to estimate motion continuously or at selected timestamps. Training combines labeled optical-flow data with self-supervision, consistency constraints, or teacher-student learning, and methods are commonly evaluated on benchmarks such as DSEC and MVSEC. A central challenge is balancing accurate, dense motion estimates and temporal consistency against limited training data, changing sensor conditions, and the latency and compute limits of real-time hardware.