CAFCNet: Cross-modality asymmetric feature complement network for RGB-T salient object detection

RGB-T salient object detection attempts to locate the most attractive target on RGB image and corresponding thermal map. Due of the intrinsic disparity between RGB and thermal modality, how to deal with the respective saliency cues becomes a key issue in RGB-T SOD. However, many current approaches try to interact multi-modal features by unidirectional interaction or symmetric interaction, neglecting their respective role in SOD task. To remedy this situation, a novel Cross-modality Asymmetric Feature Complement Network (CAFCNet) for RGB-T salient object detection is proposed. Precisely, we design a feature interaction unit called Asymmetric Feature Complement (AFC) module in order to enhance the RGB and thermal features in an asymmetric way while reducing the interference of thermal modality. Besides, we deploy a Feature Selection and Fusion (FSF) module to fuse features of two modalities in channel and spatial dimension. Finally, three Semantic Enhancement Decoder (SED) modules are used to augment and fuse features of high levels and low levels in order to build saliency map. Experiments on three different datasets show that our proposed CAFCNet has better performance compared with state-of-the-art RGB-T SOD approaches.