Title: A Scalable Event Cloud Network for Event-Based Classification

Abstract:

Biologically inspired event cameras are attracting considerable interest from both academic researchers and industry professionals. While conventional approaches typically rely on frame and voxel representations—achieving acceptable results at the cost of time-intensive conversions, model bloat, and the loss of detailed temporal data—point cloud representations offer a potential solution to these limitations. However, point clouds struggle with scalability when tasked with extracting features from high-spatial-resolution data and extended temporal event sequences.

To address these challenges, we introduce SECNet, a Scalable Network designed to utilize the Event Cloud representation. SECNet enhances performance by incorporating polarity at the structural level through a novel Event-based Group and Sampling module, moving beyond simple input-level integration. To manage the rapid increase in event volume, the network performs feature extraction in the frequency domain using the Fourier transform. This strategy significantly reduces the computational burden associated with Multiply Accumulate Operations while efficiently capturing spatio-temporal characteristics.

We validated SECNet through comprehensive testing across ten distinct event-based datasets, demonstrating its superior scalability, effectiveness, and efficiency. The source code for this work will be publicly accessible at: https://github.com/rhwxmx/SECNet_ICML.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC