This paper introduces a novel approach for Moving Object Segmentation (MOS) that utilizes Dynamic Mode Decomposition (DMD) on event streams to reconstruct missing motion dynamics between RGB frames. By integrating motion dynamics with spatial information through our Motion-Aware Segmentation Network (MAS-Net), the model requires only RGB inputs during inference, leveraging learned motion priors for event-aware performance. Our approach demonstrates robustness and effectiveness in capturing spatiotemporal dynamics in complex real-world scenarios.

The 36th British Machine Vision Conference
Events Meet Dynamic Mode Decomposition: Capturing the Spatiotemporal Dynamics of Moving Objects
Zhouning Du¹, Israr Ulhaq¹, Thanh Thi Huyen Phan¹, Yuichiro Yoshimura¹, Jigyasa Chand¹, Truong Vinh Truong Duy¹

1 Aisin Corporation

Related link:
URL: paper.pdf