New AGCN model boosts event stream recognition with dual point-voxel graphs

Event cameras generate dense asynchronous streams of pixel-level brightness changes. To process this data efficiently, researchers typically downsample events into sparse point clouds or voxel grids and then use graph neural networks (GNNs) for representation learning. However, existing methods have two shortcomings: (1) they rely on simple max or mean pooling, which ignores the relative importance of graph nodes; and (2) they use either a point or a voxel representation, missing the complementary strengths of both.

To overcome these issues, the team introduces a dual point-voxel absorbing graph convolutional network (AGCN). The model first converts event streams into both sparse point clouds and voxel grids, then builds separate absorbing graph models for each. The key innovation is the absorbing node, which acts as a global context node that learns to aggregate information from all other nodes, effectively capturing node importance and providing a more informative graph-level representation. Extensive experiments on event-based classification benchmarks (e.g., N-Caltech101, N-MNIST, ASL-DVS) show that AGCN outperforms state-of-the-art GNN-based methods by up to 5% in accuracy, demonstrating the value of fusing point and voxel representations with adaptive node weighting.