Event Voxel Set Transformer for Spatiotemporal Representation Learning on Event Streams

Event cameras are neuromorphic vision sensors representing visual information as sparse and asynchronous event streams. Most state-of-the-art event-based methods project events into dense frames and process them with conventional learning models. However, these approaches sacrifice the sparsity and high temporal resolution of event data, resulting in a large model size and high computational complexity. To fit the sparse nature of events and sufficiently explore their implicit relationship, we develop a novel attention-aware framework named Event Voxel Set Transformer (EVSTr) for spatiotemporal representation learning on event streams. It first converts the event stream into a voxel set and then hierarchically aggregates voxel features to obtain robust representations. The core of EVSTr is an event voxel transformer encoder to extract discriminative spatiotemporal features, which consists of two well-designed components, including a multi-scale neighbor embedding layer (MNEL) for local information aggregation and a voxel self-attention layer (VSAL) for global representation modeling. Enabling the framework to incorporate a long-term temporal structure, we introduce a segmental consensus strategy for modeling motion patterns over a sequence of segmented voxel sets. We evaluate the proposed framework on two event-based tasks: object classification and action recognition. Comprehensive experiments show that EVSTr achieves state-of-the-art performance while maintaining low model complexity. Additionally, we present a new dataset (NeuroHAR) recorded in challenging visual scenarios to address the lack of real-world event-based datasets for action recognition.

A Dynamic Graph CNN with Cross-Representation Distillation for Event-Based Recognition

It is a popular solution to convert events into dense frame-based representations to use the well-pretrained CNNs in hand. Although with appealing performance, this line of work sacrifices the sparsity/temporal precision of events and usually necessitates heavy-weight models, thereby largely weakening the advantages and real-life application potential of event cameras. A more application-friendly way is to design deep graph models for learning sparse point-based representations from events. Yet, the efficacy of these graph models is far behind the frame-based counterpart with two key limitations: ($i$) simple graph construction strategies without carefully integrating the variant attributes (i.e., semantics, spatial and temporal coordinates) for each vertex, leading to biased graph representation; ($ii$) deficient learning because the lack of well pretraining models available. Here we solve the first problem by introducing a new event-based graph CNN (EDGCN), with a dynamic aggregation module to integrate all attributes of vertices adaptively. To alleviate the learning difficulty, we propose to leverage the dense representation counterpart of events as a cross-representation auxiliary to supply additional supervision and prior knowledge for the event graph. To this end, we form a frame-to-graph transfer learning framework with a customized hybrid distillation loss to well respect the varying cross-representation gaps across layers. Extensive experiments on multiple vision tasks validate the effectiveness and high generalization ability of our proposed model and distillation strategy (Core components of our codes are submitted with supplementary material and will be made publicly available upon acceptance)