Event-camera optical flow reaches new efficiency frontiers with mesh-based representations

Learning Efficient Meshflow and Optical Flow from Event Cameras (arxiv 2510.04111) proposes a mesh-based representation for event-camera scene flow estimation that achieves comparable accuracy to dense optical flow at substantially lower computational cost.

The optical flow density tradeoff

Standard optical flow estimates motion at every pixel in the image. For an event camera, this means assigning a velocity to every possible pixel location — including pixels that are generating no events because the scene at that location is static. Computing a dense flow field from sparse event data wastes computation at every pixel where no events arrived.

The mesh-based approach represents flow using a sparse mesh of control points whose positions are learned from the event data. The flow at any arbitrary point is interpolated from the nearest control points. For a scene where 90% of pixels are static (as is typical for a camera with a distant fast-moving target), the mesh needs only a handful of control points near the target, while the background can be represented with a very coarse mesh.

Efficiency gains

The paper reports 4–6x reduction in computation time compared to dense optical flow methods at comparable accuracy on standard benchmarks, using a mesh density that adapts to the local event density. High-event regions (the target, moving objects) get a fine mesh; low-event regions (static background) get a coarse mesh.

For a processing pipeline that must run at event rates on embedded hardware, this efficiency gain directly translates to a lower compute requirement or a higher event rate that can be processed in real time.

Accuracy comparison

Accuracy on MVSEC and DSEC is comparable to dense flow methods, with the mesh-based approach slightly underperforming on scenes with many independently-moving objects (where many mesh regions need fine control) and performing comparably or better on scenes with a single fast-moving target against a static background.

The single-target-against-static-background scenario is precisely the terminal seeker scenario — the target is the only moving thing in the seeker’s field of view during the terminal phase, and the background is the sky. Meshflow’s efficiency advantage is largest in this scenario.

Implementation path

Mesh-based representations are computationally simpler than dense flow representations and have known efficient hardware implementations. Control point grids with local interpolation are a standard operation in texture mapping and graphics hardware. The efficient meshflow architecture described in this paper has a clearer path to FPGA implementation than dense attention-based flow architectures.

For the seeker design problem, this matters: the FPGA fabric available in an embedded package is limited, and algorithms that can be mapped to regular, local operations on sparse data are much easier to implement efficiently than algorithms with global dependencies.

---

Learning Efficient Meshflow and Optical Flow from Event Cameras. arxiv 2510.04111