SPATIOTEMPORAL REGISTRATION FOR EVENT-BASED VISUAL ODOMETRY

Event: CVPR 2025 · Duration: 5 min · ▶ Watch on YouTube

Abstract

Event cameras offer significant advantages over traditional RGB cameras, including high temporal resolution, low latency, and sensitivity to light changes, making them highly suitable for visual odometry. Existing event-based motion estimation methods, such as Contrast Maximization, often face challenges with real-time performance or maintaining accuracy across diverse operating conditions. This work proposes a novel Spatiotemporal Registration (STR) framework that directly estimates motion parameters by warping event streams and minimizing reprojection error. The proposed method demonstrates robustness to variations in batch size and consistently outperforms existing techniques in both accuracy and runtime, enabling real-time operation for event-based visual odometry.

Speakers

  • Daqi Liu — Australian Institute for Machine Learning, The University of Adelaide
  • Alvaro Parra — Australian Institute for Machine Learning, The University of Adelaide
  • Tat-Jun Chin — Australian Institute for Machine Learning, The University of Adelaide

Talks (1)

  • 00:00:00 — Daqi Liu: SPATIOTEMPORAL REGISTRATION FOR EVENT-BASED VISUAL ODOMETRY
    • This paper introduces a novel spatiotemporal registration framework for event-based visual odometry that achieves high accuracy and real-time performance by leveraging the unique properties of event cameras and an EM-type optimization approach.

Key Takeaways

  • Event cameras provide asynchronous, high temporal resolution, low latency, and high HDR information, making them well-suited for visual odometry tasks.
  • Existing event-based motion estimation methods, such as Contrast Maximization, often struggle to achieve real-time performance.
  • The proposed Spatiotemporal Registration (STR) framework utilizes a novel warping function and an EM-type optimization to achieve accurate and real-time motion estimation.
  • STR generates feature tracks as a byproduct, which can be effectively used for rotation averaging or pose graph optimization.
  • Quantitative results demonstrate that STR outperforms Contrast Maximization and other state-of-the-art methods in both accuracy and runtime, particularly under varying batch sizes and challenging low-light conditions, processing up to 1 million events per second.

Methods / Models / Datasets Mentioned

  • Contrast Maximization (CM)
  • Global optimal Contrast Maximization
  • Reward function for Contrast maximization
  • Entropy Minimization
  • Spatiotemporal Registration (STR)
  • VCM
  • VEM
  • ZHU
  • ROBOTEVT
  • iniVation 240C event camera
  • UR-5 robot arm
  • PureRot
  • ParRot
  • PureTranslate
  • FullMod

Topics

Event-based visual odometry · Spatiotemporal registration · Motion estimation · Event cameras · Contrast Maximization · Real-time performance · Feature tracking · Dataset collection

Notes

Open for commentary — connections to other work, critiques, follow-up reading.