RGB-D Salient Object Detection with Ubiquitous Target Awareness

Conventional RGB-D salient object detection methods aim to leverage depth as complementary information find the regions in both modalities. However, results heavily rely on quality of captured data which sometimes are unavailable. In this work, we make first attempt solve problem with a novel depth-awareness framework. This framework only relies RGB testing phase, utilizing supervision for representation learning. To construct our well achieving accurate results, propose Ubiquitous Target Awareness (UTA) network three important challenges SOD task: 1) awareness module excavate and mine ambiguous via adaptive depth-error weights, 2) spatial-aware cross-modal interaction channel-aware cross-level interaction, exploiting low-level boundary cues amplifying high-level channels, 3) gated multi-scale predictor perceive saliency different contextual scales. Besides its high performance, proposed UTA is depth-free inference runs real-time 43 FPS. Experimental evidence demonstrates that not surpasses state-of-the-art five public benchmarks by large margin, but also verifies extensibility benchmarks.