DIMOS: Disentangling Instance-level Moving Object Segmentation

The paper introduces DIMOS, a novel framework for moving instance segmentation (MIS) that addresses the challenges of segmenting small moving objects using event cameras. It features a dual-disentangling extraction method to separate appearance and motion information, coupled with a multi-granularity cross-modal alignment for improved feature fusion. The proposed approach achieves state-of-the-art performance in MIS benchmarks, particularly in scenarios involving fast motion and low-light conditions, making it significant for applications in autonomous driving and surveillance.