Deep visual odometry for stereo event cameras reaches deployment-level accuracy

Researchers demonstrate learned visual odometry from stereo event cameras that achieves positioning accuracy competitive with state-of-the-art frame-based methods — at update rates an order of magnitude higher.

Deep Visual Odometry for Stereo Event Cameras (arxiv 2509.08235) demonstrates a complete learned odometry pipeline that takes stereo event streams as input and produces six-degree-of-freedom pose estimates at rates that frame-based approaches cannot match. The results are significant not just for mobile robotics — the stated application domain — but for any system requiring high-update-rate state estimation on a moving platform.

The stereo event advantage

Monocular event cameras can estimate rotation well but scale-ambiguous translation poorly. Stereo configurations solve the scale problem by triangulating depth directly, but the standard stereo processing pipeline — disparity estimation, 3D feature triangulation, bundle adjustment — was designed for frame cameras and introduces latency that undermines the event camera’s speed advantage.

This paper builds a stereo pipeline from scratch for events. Rather than adapting frame-camera algorithms, it trains a network end-to-end on stereo event streams and 6-DOF ground truth, allowing the network to learn disparity-consistent representations without the intermediate disparity estimation step.

Update rate and latency

The most practically significant result is update rate: their pipeline produces a pose estimate every 1–5 milliseconds, depending on event rate. Comparable frame-based visual odometry running at 30 fps cannot produce a new estimate faster than 33 milliseconds, and higher-speed frame systems at 240 fps require pixel counts that create downstream bandwidth problems.

For a moving platform that needs to maintain attitude and position knowledge during a short, high-rate maneuver — a countermeasure eject, a seeker-head stabilization loop, a precision guidance correction — the 1-5 ms update rate matters. A 33 ms update rate misses an entire cycle of many guidance loops.

Accuracy compared to frame-based methods

The accuracy comparison shows stereo event visual odometry achieving translation errors below 1% of distance traveled on their benchmark dataset. This is competitive with frame-based methods on the same sequences, which is notable because frame-based methods are more mature and have had more development effort.

In high-angular-rate scenarios — which are explicitly evaluated in this dataset — the event camera significantly outperforms frame cameras. In slow, smooth motion the margin narrows. The event camera earns its advantage precisely in the conditions where the guidance problem is hardest.

Platform considerations

The paper notes that stereo event camera baseline distance matters: a narrow stereo baseline limits depth accuracy at range, while a wide baseline increases the size and mass of the sensor head. For an interceptor seeker package, this is a real integration tradeoff.

Their results suggest that a baseline of roughly 10 cm produces useful depth estimates for targets at 10–50 meter range — the terminal engagement distance — but the geometry becomes marginal much beyond 50 meters. For terminal-phase sensing this is acceptable; for long-range initial detection it is not, which suggests stereo event cameras as one element of a layered sensing chain rather than a standalone solution.

Deep Visual Odometry for Stereo Event Cameras. arxiv 2509.08235