Research map / 3D vision / Depth and stereo

E-MoFlow: Learning Egomotion and Optical Flow from Event Data via Implicit Regularization

NeurIPS 2025 · 3D vision › Depth and stereo · approach: Learned optical and scene flow · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Unsupervised joint estimation of event-camera optical flow (neural implicit flow field) and 6-DoF egomotion (cubic B-spline) using a contrast-maximization loss plus a depth-free differential epipolar geometric constraint.

Problem

Estimating optical flow and 6-DoF egomotion jointly from event camera data is ill-posed due to lack of long-term data association; prior methods either add explicit smoothness regularizers that bias results or add depth parameters that cause local minima.

Approach

Model flow as a coordinate-based MLP (implicit neural representation) integrated via neural ODE for event warping, model egomotion as a cubic B-spline, and couple them through a depth-free differential epipolar constraint so that implicit smoothness and geometric consistency regularize the solution without explicit depth or smoothness terms.

Key results

• On MVSEC optical flow (dt=1), E-MoFlow achieves average EPE 0.450 and %Out 0.328, best among unsupervised methods and second overall behind MultiCM-V2 (0.348 EPE).
• On MVSEC 6-DoF egomotion without depth priors, average RMS angular error 4.062 and linear error 0.285, versus 8.900 and 1.680 for MultiCM-V2.
• On DSEC, E-MoFlow performs comparably to unsupervised baselines (average EPE 3.14) but trails supervised methods such as TMA (0.74).

Measured effects reported in the paper

Surprising findings

• The method's 6-DoF estimation is reported to match ground-truth-motion supervision on some MVSEC sequences (e.g., indoor_flying2 and indoor_flying3 at dt=4).
• Early stopping gives 2.2x training speedup at only about 2-3% EPE cost.

Limitations

• Performance on DSEC lags supervised methods, attributed to complex driving motion, illumination changes and unstructured depth.
• May produce errors at edges in scenes with complex textures and drastic depth changes.
• Per-segment optimization (1k iterations per short event segment) rather than a feedforward model.

Experimental setup

Datasets: MVSEC DSEC

Benchmarks: MVSEC DSEC

Metrics: EPE AE %Out RMSE angular velocity RMSE linear velocity

Baselines: MultiCM-V2 MotionPriorCMax Paredes et al. EV-FlowNet E-RAFT TMA ADM-Flow AEmin PEME

Builds on: contrast maximization neural ODE NeRF differential epipolar constraint

Closest papers by meaning

• EgoFlow: Gradient-Guided Flow Matching for Egocentric 6DoF Object Motion Generation (CVPR 2026 · similarity 0.91)
• Flow-NeRF: Joint Learning of Geometry, Poses, and Dense Flow within Unified Neural Representations (CVPR 2025 · similarity 0.87)
• Full-DoF Egomotion Estimation for Event Cameras Using Geometric Solvers (CVPR 2025 · similarity 0.86)
• FlowFeat: Pixel-Dense Embedding of Motion Profiles (NeurIPS 2025 · similarity 0.86)
• H-Flow: Self-supervised Human Scene Flow via Physics-inspired Joint Multi-modal Learning (NeurIPS 2026 · similarity 0.86)
• EDCFlow: Exploring Temporally Dense Difference Maps for Event-based Optical Flow Estimation (CVPR 2025 · similarity 0.86)