A comprehensive survey of event-based stereo depth estimation maps where the field stands

Depth from stereo requires solving the correspondence problem — finding the same scene point in two images taken from different positions. For frame cameras this is well-understood, with decades of algorithms and hardware support. For event cameras the problem is harder and more interesting: the two sensors produce asynchronous, sparse event streams, and correspondence now means finding matching events across both streams in time as well as space.

A survey paper on arxiv (2409.17680) covers this field systematically, from the early methods that treated stereo events like stereo frames — binning events into time windows and running standard disparity algorithms — through to methods that work directly on the asynchronous event stream. Reading it prompted us to think through what stereo depth from events would actually mean for airborne seeker applications.

Methods taxonomy

The survey organizes approaches into three main families. The first bins events into frames or voxel grids and applies adapted 2D stereo algorithms; these inherit the temporal resolution limitations of the representation but leverage mature stereo infrastructure. The second operates on event graphs or point clouds directly, exploiting the temporal precision of the raw data; these are more computationally demanding but preserve the sensor’s native advantages. The third uses continuous-time formulations — typically based on time surfaces or exponential kernels — that represent the most recent event at each pixel as a decaying function of time; correspondence is then found in this continuous-time representation rather than in discrete bins.

The performance benchmarks in the survey make clear that the continuous-time and graph-based methods consistently outperform the frame-based methods on fast-motion scenes, which is exactly the regime that matters for terminal guidance. The cost is compute: the frame-based pipeline can run on modest embedded hardware, while the continuous-time approaches typically need GPU-class resources in current implementations.

What stereo depth adds for seekers

A monocular event-camera seeker can track a target’s angular position but not its range. Range estimation from monocular motion is possible — the relationship between target size and range is known for a given target class — but this requires assumptions about target geometry that may not hold for cooperative-looking targets that have been modified to minimize signature.

Stereo depth from a calibrated event-camera pair gives range directly, independent of target size assumptions. For a terminal seeker, range is an input to the proportional navigation law and to the fuzing condition. Having direct range measurement rather than relying on scene geometry or time-to-intercept estimates improves guidance performance and reduces sensitivity to target size uncertainty.

The engineering constraint is baseline. The stereo depth precision scales with the baseline divided by the range; at 200 meters a 10 cm baseline gives centimeter-level precision, which is sufficient for guidance. At 50 meters the precision improves proportionally. The question is whether a 10 cm stereo rig can be packaged into an interceptor seeker without exceeding the volume budget, and whether the calibration remains stable under the vibration and temperature environment of flight.

Open problems identified in the survey

The survey identifies several open problems directly relevant to our use case. Calibration between stereo event cameras is more demanding than calibration between stereo frame cameras because the calibration target needs to produce events — static targets don’t. Moving calibration targets introduce their own uncertainty. Second, handling large velocity gradients — a target that is fast in one region of the frame and slow in another — remains difficult for all current methods. Third, there are no publicly available stereo event datasets collected from an airborne platform tracking aerial targets; all existing benchmarks are ground-based.

The absence of a benchmark dataset in the relevant operating regime is actually an opportunity. Collecting stereo event data during flight tests produces the raw material for a dataset that would be directly relevant to the terminal guidance problem and doesn’t exist elsewhere. That data has value beyond internal development.