Cross-modal co-attention network for light field salient object detection

Light-field salient object detection aims to segment the visually distinctive objects out of surroundings. Due to the fact that light field images provide all-in-focus images and focal stacks for the same scene, they record comprehensive but redundant information. Meanwhile, the noticeable modal differences between RGB images and focal stack make cross-modal fusion a key challenge. To address these issues, we propose a cross-modal co-attention network (CCNet). First, weights are assigned to focal slices based on their importance, with higher weights assigned to slices containing clear objects to maximize the selection of valuable information. Then, a cross-complementary refinement (CCR) module and a cross-modal hierarchical co-attention (CHC) module are designed to achieve effective fusion of RGB features and focal stack features while fully considering the information exchange between layered features. Finally, we propose a RGB-guided decoding module to gradually decode and generate clear saliency maps under the guidance of RGB modality. Extensive experimental results demonstrate that CCNet achieves superior performance against the state-of-the-art methods across multiple benchmark datasets.