Feature aggregation with transformer for RGB-T salient object detection

Abstract

The main purpose of RGB-T salient object detection (SOD) is to fully integrate and exploit the information from the complementary fusion of modalities to address the underperformance of RGB SOD in some challenging scenes. In this paper, we propose a novel feature aggregation network that can fully mine multi-scale and multi-modal information for complete and accurate RGB-T SOD. Subsequently, a cross-attention fusion module is proposed to adaptively integrate high-level features by using the attention mechanism in the Transformer. Then we design a simple yet effective fast feature aggregation module to fuse low-level features. Through the combined work of the above modules, our network can perform well in some complex scenes by effectively fusing features from RGB and thermal modalities. Finally, several experiments on publicly available datasets such as VT821, VT1000, and VT5000 demonstrate that our method outperforms state-of-the-art methods. And our code has been released at:https://github.com/ELOESZHANG/FANet.