Point-voxel absorbing graph representation learning for event stream-based recognition

Abstract

Sampled point- and voxel-based representations are commonly used to compress dense event streams for efficient graph-based processing, where points or voxels are treated as graph nodes and encoded using graph neural networks (GNNs). However, existing event-based GNNs suffer from two main limitations. (1) they typically employ max or mean pooling to aggregate node embeddings into a graph-level representation, which overlooks differences in node importance and limits global information interaction. (2) most methods rely on either pointbased or voxel-based graphs, without exploiting the complementary characteristics of the two representations. To address these limitations, we propose a dual-branch point-voxel absorbing graph framework for event-stream recognition. Specifically, an input event stream is transformed into a sparse event cloud and a set of voxel grids, from which two absorbing graphs are constructed. We further develop an absorbing graph convolutional network (AGCN), in which an absorbing node interacts with all event nodes to perform node-aware global information aggregation. The graph-level representations learned by the point and voxel branches are then concatenated and fed into an MLP classifier. Extensive experiments on multiple event-based classification benchmarks demonstrate the effectiveness of the proposed framework.