Inertia-Informed Orientation Priors for Event-Based Optical Flow Estimation

TL;DR

This work tackles the non-convexity and initialization sensitivity of contrast-maximization (CM) for event-based optical flow by introducing inertia-informed orientation priors derived from 3D camera velocities. By generating orientation maps and integrating them as priors into a hybrid CM objective, the method improves robustness and convergence, while remaining modular to fall back when egomotion cues are unavailable. The approach, evaluated on MVSEC, DSEC, and ECD, achieves superior or competitive accuracy relative to state-of-the-art model-based methods and demonstrates clear benefits in low-texture or challenging motion scenarios. The work highlights the practical impact of incorporating visual–vestibular-inspired priors to enhance real-time event-based perception in robotics and autonomous systems, with avenues for richer priors in future extensions.

Abstract

Event cameras, by virtue of their working principle, directly encode motion within a scene. Many learning-based and model-based methods exist that estimate event-based optical flow, however the temporally dense yet spatially sparse nature of events poses significant challenges. To address these issues, contrast maximization (CM) is a prominent model-based optimization methodology that estimates the motion trajectories of events within an event volume by optimally warping them. Since its introduction, the CM framework has undergone a series of refinements by the computer vision community. Nonetheless, it remains a highly non-convex optimization problem. In this paper, we introduce a novel biologically-inspired hybrid CM method for event-based optical flow estimation that couples visual and inertial motion cues. Concretely, we propose the use of orientation maps, derived from camera 3D velocities, as priors to guide the CM process. The orientation maps provide directional guidance and constrain the space of estimated motion trajectories. We show that this orientation-guided formulation leads to improved robustness and convergence in event-based optical flow estimation. The evaluation of our approach on the MVSEC, DSEC, and ECD datasets yields superior accuracy scores over the state of the art.

Figures (9)

• Figure 1: The camera motion and image plane geometry (left) and induced orientation maps and flow patterns (right) due to camera linear (yellow) and angular (purple) velocities. Structured inertia-informed orientation priors generated in the form of orientation maps from camera motion cues are used to provide guidance to the CM process for event-based optical flow estimation.
• Figure 2: Singularity points due to camera velocity. The linear motion of a 1D camera is depicted along a parabola with instantaneous velocity vectors (yellow). This velocity ray intersects with the image plane (in- or out-of-sensor) at the singularity point at which zero optical flow is observed. The corresponding color coded 1D flow orientations are illustrated for each pose, where cyan and red colors indicate left and right flow orientations, respectively. The top-left and top-right 1D arrays demonstrate the FOC and FOE, respectively.
• Figure 3: Generating distorted orientation maps. (a) and (b) depict the camera linear and angular velocities along with their corresponding intersections $\boldsymbol{s}_{\text{lin}}$ and $\boldsymbol{s}_{\text{ang}}$ with the image plane, respectively. (c) and (d) demonstrate the final generated linear and angular orientation maps corresponding to (a) and (b), respectively.
• Figure 4: Applying lens distortion to the orientation maps. (a) shows an example of an undistorted orientation map. (b) illustrates the distorted map with empty regions as a result of applying lens distortion. (c) and (d) depict the filled map and its difference with the undistorted map, respectively.
• Figure 5: Orientation priors for CM workflow diagram.