HRTransNet: HRFormer-Driven Two-Modality Salient Object Detection

The High-Resolution Transformer (HRFormer) can maintain high-resolution representation and share global receptive fields. It is friendly towards salient object detection (SOD) in which the input output have same resolution. However, two critical problems need to be solved for two-modality SOD. One problem fusion. other HRFormer output's To address first problem, a supplementary modality injected into primary by using optimization an attention mechanism select purify at level. solve second dual-direction short connection fusion module used optimize features of HRFormer, thereby enhancing detailed objects proposed model, named HRTransNet, introduces auxiliary stream feature extraction modality. Then, are beginning each multi-resolution branch. Next, applied achieve forwarding propagation. Finally, all with different resolutions aggregated intra-feature inter-feature interactive transformers. Application model results impressive improvement driving SOD tasks, e.g., RGB-D, RGB-T, light field SOD.https://github.com/liuzywen/HRTransNet