MVT: Multi-Vision Transformer for Event-Based Small Target Detection

Abstract

Object detection in remote sensing plays a crucial role in various ground identification tasks. However, due to the limited feature information contained within small targets, which are more susceptible to being buried by complex backgrounds, especially in extreme environments (e.g., low-light, motion-blur scenes). Meanwhile, event cameras offer a unique paradigm with high temporal resolution and wide dynamic range for object detection. These advantages enable event cameras without being limited by the intensity of light, to perform better in challenging conditions compared to traditional cameras. In this work, we introduce the Multi-Vision Transformer (MVT), which comprises three efficiently designed components: the downsampling module, the Channel Spatial Attention (CSA) module, and the Global Spatial Attention (GSA) module. This architecture simultaneously considers short-term and long-term dependencies in semantic information, resulting in improved performance for small object detection. Additionally, we propose Cross Deformable Attention (CDA), which progressively fuses high-level and low-level features instead of considering all scales at each layer, thereby reducing the computational complexity of multi-scale features. Nevertheless, due to the scarcity of event camera remote sensing datasets, we provide the Event Object Detection (EOD) dataset, which is the first dataset that includes various extreme scenarios specifically introduced for remote sensing using event cameras. Moreover, we conducted experiments on the EOD dataset and two typical unmanned aerial vehicle remote sensing datasets (VisDrone2019 and UAVDT Dataset). The comprehensive results demonstrate that the proposed MVT-Net achieves a promising and competitive performance.