Salient Object Detection

Salient object detection predicts pixel-level masks for regions that stand out, helping vision systems focus on foreground objects for segmentation, retrieval, and scene understanding. Methods combine multi-scale CNN or transformer features with boundary refinement; newer approaches also use state-space models, diffusion-based mask refinement, and pretrained vision or segmentation models, while RGB-D, RGB-thermal, video, hyperspectral, remote-sensing, and panoramic settings add complementary cues. Systems are evaluated on mask benchmarks such as DUTS and ECSSD and modality-specific RGB-D and video datasets, using measures including F-measure, MAE, and structural similarity; training ranges from fully supervised to weakly supervised, self-supervised, and training-free approaches. A central challenge is knowing when extra modalities or foundation-model features are reliable and aligned while preserving boundaries and small objects, and benchmarks often assume visual prominence even when saliency depends on task or user intent.