MPM-Net: A Network with Multiple Pluggable Modules for RGB-D Salient Object Detection
Mingyuan Lei, Shangkun Wen, Qinglin Sun · 2025
RGB-D salient object detection, combining RGB visual appearance and depth cues, aims to imitate the human visual system's attention mechanism and locate the most prominent objects or regions in a scene. However, existing models struggle to balance parameter quantity and accuracy, and lack flexibility. This paper introduces MPM-Net, leveraging a series of highly flexible and adaptable pluggable modules. The Local Guidance Information Access (LGIA) module, upon gathering fused feature images of various scales, enables more comprehensive cross-modality interactions. The Cross-Compensation Module (CCM) compensates the detection results from the RGB, depth, and RGB-D streams each other, effectively improving the model's accuracy. MPM-Net can select modules according to computing resources, freely adjusting the parameter-accuracy relationship for different tasks. Its strong flexibility offers powerful support for applying deep learning salient object detection in diverse fields and scenarios.