Appearance-Boundary Fusion Distillation for Event-Based Optical Flow

Abstract

Event cameras can perceive the continuous motion of a scene at high temporal resolution, thereby providing effective motion cues for optical flow estimation. In recent years, some methods have attempted to jointly leverage images and events to improve performance. However, most of them directly fuse the two modalities, overlooking the differences between image appearance and event boundaries, which easily leads to modal interference and limits the utilization of complementary information. Moreover, to meet the practical deployment requirement of event-only input, existing distillation methods still struggle to effectively handle the heterogeneous modality gap between images and events. To this end, we propose an Appearance–Boundary Fusion Distillation framework (ABFD) for event-based optical flow estimation. Specifically, the proposed method first employs an Appearance–Boundary Decoupled Fusion module (ABF) to explicitly model image appearance information and event boundary information, and further generates high-quality fused features through asymmetric cross-modal interaction. Subsequently, a Motion-Aware Distillation strategy (MAD) is introduced to transfer effective motion representations from the fused teacher to the event-only student model under the supervision of the optical flow task. During inference, only the event-based model is retained, enabling efficient and robust optical flow estimation without requiring additional image input. Extensive experiments demonstrate that the proposed method achieves superior performance over existing methods.