Salient object detection wants a clean mask of whatever stands out in an image, and doing it with RGB plus depth usually means painstaking pixel labels. Scribble labels, a few careless strokes per object, are far cheaper but too sparse to pin down edges and shape, so the maps come out mushy.

The fix here has two parts. First, SAM-PAG uses the Segment Anything Model to grow those scribbles into dense pixel-level pseudo labels, with a dual-branch design and mask consistency keeping them honest. Second, S2Diff is a conditional diffusion model that starts from a noisy saliency map and denoises toward a sharp one, guided by cross-modal features built through frequency integration and a state-space interaction between the color and depth streams. Across seven datasets it beats other scribble-supervised methods and stays competitive with fully supervised ones.

Plenty of moving parts, so the paper and code are where they fit together.