O-MMS: Zero-Shot Multi-Motion Segmentation With A Monocular Event Camera

Event: CVPR 2025 · Duration: 2 min · ▶ Watch on YouTube

Abstract

Nature relies on motion detection for fast obstacle avoidance, a task difficult for classical cameras due to motion blur. This work proposes a hybrid solution for zero-shot multi-motion segmentation using monocular event cameras, which capture asynchronous intensity changes, enabling high-speed computation. The method combines bottom-up feature tracking with top-down model fitting, and introduces a new open-source benchmark dataset, MOD++, to facilitate systematic evaluation of event-based motion segmentation algorithms across various motion types.

Speakers

  • Chethan M. Parameshwara — University of Maryland
  • Nitin J. Sanket — University of Maryland
  • Chahat Deep Singh — University of Maryland
  • Cornelia Fermüller — University of Maryland
  • Yiannis Aloimonos — University of Maryland

Talks (1)

  • 00:00:00 — Chethan M. Parameshwara, Nitin J. Sanket, Chahat Deep Singh, Cornelia Fermüller, Yiannis Aloimonos: O-MMS: Zero-Shot Multi-Motion Segmentation With A Monocular Event Camera
    • This paper presents a hybrid solution for zero-shot multi-motion segmentation using a monocular event camera, combining bottom-up feature tracking and top-down model fitting, and introduces a new open-source benchmark dataset called MOD++.

Key Takeaways

  • Event cameras are superior to classical cameras for motion detection due to their ability to capture asynchronous intensity changes without motion blur, enabling faster reaction times.
  • The proposed hybrid approach combines bottom-up feature tracking and top-down model fitting for robust zero-shot multi-motion segmentation.
  • A novel open-source benchmark dataset, MOD++, is introduced to address the lack of stratified data for evaluating event-based motion segmentation under diverse motion conditions.
  • The method successfully segments independently moving objects in complex real-world and synthetic scenarios, including challenging sequences with breaking objects.

Methods / Models / Datasets Mentioned

  • DVS (Dynamic Vision Sensor)
  • Feature tracklets
  • MOD++
  • EV-IMO
  • EED

Topics

Zero-shot multi-motion segmentation · Monocular event camera · Motion detection · Event cameras · Feature tracking · Model fitting · MOD++ dataset · High-speed computation · Obstacle avoidance

Notes

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