In RGB-D salient object detection the colour image supplies semantic content and the depth map supplies spatial-geometric structure, and which one dominates changes from scene to scene - so fusing them on a fixed hand-specified schedule wastes the complementarity. Mutual fusion instead lets each single-modality feature interact with the fused feature itself, once along the spatial (position) axis and once along the channel axis, so the weighting is learned per scene rather than fixed. Learners work through the position and channel mutual fusion computations, including the pooled spatial-attention map and the normalizations that keep the matrix products stable, and then through a decoder that applies an independent, unshared transformer stage to each fusion level so global context is added without the cost of repeatedly running one shared transformer. The lesson generalises to any two-modality dense-prediction problem where the modalities disagree about what matters.
GL-DMNet is a network for RGB-D salient object detection - the pixel-level task of producing a binary map of the most visually prominent object in a scene, given a colour image and its aligned depth …