Revisiting Feature Fusion for RGB-T Salient Object Detection

While many RGB-based saliency detection algorithms have recently shown the capability of segmenting salient objects from an image, they still suffer unsatisfactory performance when dealing with complex scenarios, insufficient illumination or occluded appearances. To overcome this problem, article studies RGB-T detection, where we take advantage thermal modality's robustness against and occlusion. achieve goal, revisit feature fusion for mining intrinsic patterns propose a novel deep network, which consists multi-scale, multi-modality, multi-level modules. Specifically, multi-scale module captures rich contexture features each modality feature, while multi-modality modules integrate complementary different level features, respectively. demonstrate effectiveness proposed approach, conduct comprehensive experiments on benchmark. The experimental results that our approach outperforms other state-of-the-art methods conventional by large margin.