Cross-space Attentive Fusion of Hyperbolic and Euclidean Graph Convolutional Network for Event-based Object Detection

Event streams, which encode brightness variations asynchronously and sparsely, provide unique benefits for event-based visual perception and neural representation learning, making them particularly suitable for tasks that require fine-grained temporal modeling, such as object detection. While traditional Graph Neural Networks (GNNs) based on neighborhood aggregation and message passing excel at characterizing local structures, they suffer from a geometric mismatch with the inherent physical-geometric hierarchy of event streams, which limits their ability to capture global semantic dependencies and spatio-temporal relationships. To this end, we propose an Event-based Cross-space Attentive Fusion Graph Convolutional Network (ECAF-GCN) that leverages a dual geometry-aware architecture in Euclidean and hyperbolic spaces, respectively enabling targeted enhancement of structural and hierarchical features. To dynamically balance and fuse information across geometries, we introduce cross-space fusion to effectively and efficiently alleviate geometric mismatch and enhance hierarchical modeling for hybrid event perception. Extensive experiments validate the effectiveness of our approach. Our code will be released for public use.