Research map / 3D vision / Depth and stereo
Full-DoF Egomotion Estimation for Event Cameras Using Geometric Solvers
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.
Proposes sparse geometric solvers that estimate full-DoF event camera egomotion (angular and linear velocity) from line-induced eventail manifolds, using an incidence or a new normal-vector coplanarity formulation optimized with Adam.
Problem
Sparse geometric egomotion estimation for event cameras has only recovered linear velocity assuming known rotation (e.g., from an IMU); recovering full-DoF angular and linear velocities from sparse events without extra sensors or priors has not been explored.
Approach
Two formulations on eventail manifolds of line events: an incidence-relation eigenvalue-minimization over angular velocity (IncMin/IncBat) and a novel coplanarity relation among event-induced plane normals from normal flow (CopMin/CopBat), with a first-order rotation approximation enabling data compression and a cascade of approximate then exact rotation optimized by Adam.
Key results
- On noise-free synthetic data (M=5, N=100), the CopBat+cascade solver reaches median angular error 9.3e-5 and runtime 17.1 ms, while IncBat+cascade reaches 1.6e-4 with 48.6 ms.
- On real VECtor sequences, IncBat and CopBat achieve median angular velocity errors of about 0.20-0.24 and linear velocity errors of 17-25 degrees.
- Runtimes of the proposed solvers range from 16.7 to 48.7 ms on a laptop CPU.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use cascade rotation parameterization instead of first-order approximated rotation (ablation) | IncBat · Synthetic, noise-free (M = 5, N = 100) | median εang: 1.4 × 10−2 -> 1.6 × 10−4 | Helped |
| Use cascade rotation parameterization instead of exact rotation only (ablation) | IncBat · Synthetic, noise-free (M = 5, N = 100) | runtime: 48.7 ms -> 48.6 ms | Helped |
| Use cascade rotation parameterization instead of first-order approximated rotation (ablation) | CopBat · Synthetic, noise-free (M = 5, N = 100) | median εang: 1.9 × 10−2 -> 9.3 × 10−5 | Helped |
| Use cascade rotation parameterization instead of exact rotation only (ablation) | CopBat · Synthetic, noise-free (M = 5, N = 100) | runtime: 32.7 ms -> 17.1 ms | Helped |
| Use incidence rather than coplanarity formulation | Full-DoF egomotion estimation · desk-normal | median εang: 0.236 -> 0.232 | Helped |
| Use incidence rather than coplanarity formulation | Full-DoF egomotion estimation · sofa-normal | median εang: 0.221 -> 0.229 | Hurt |
| Use incidence rather than coplanarity formulation | Full-DoF egomotion estimation · mountain-normal | median εang: 0.216 -> 0.195 | Helped |
Surprising findings
- The first-order approximated rotation alone yields considerably lower accuracy, but cascading it with exact rotation optimization gives the best accuracy and efficiency.
Limitations
- Real-world errors are relatively large (angular velocity error about 0.2, linear velocity error about 20 degrees).
- Pure rotation cases are not handled by the previous linear solver and require a separate treatment described only in the supplement.
- Requires locally constant velocity over a short time interval and normal flow estimation from time surfaces.
Experimental setup
Datasets: VECtor dataset
Benchmarks: VECtor dataset (desk-normal, mountain-normal, sofa-normal)
Metrics: angular velocity error (normalized) linear velocity angular error (degrees) success rate runtime
Baselines: IncBat CopBat
Builds on: eventail manifold Adam contrast maximization
Closest papers by meaning
- An N-Point Linear Solver for Line and Motion Estimation with Event Cameras (CVPR 2024 · similarity 0.88)
- E-MoFlow: Learning Egomotion and Optical Flow from Event Data via Implicit Regularization (NeurIPS 2025 · similarity 0.86)
- Scalable Benchmarking and Robust Learning for Noise-Free Ego-Motion and 3D Reconstruction from Noisy Video (ICLR 2025 · similarity 0.83)
- Event-based Structure-from-Orbit (CVPR 2024 · similarity 0.83)
- Efficient Meshflow and Optical Flow Estimation from Event Cameras (CVPR 2024 · similarity 0.83)
- EF-3DGS: Event-Aided Free-Trajectory 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.