Cascade Graph Neural Networks for RGB-D Salient Object Detection

Abstract

In this paper, we study the problem of salient object detection for RGB-D images by using both color and depth information. A major technical challenge for detecting salient objects in RGB-D images is to fully leverage the two complementary data sources. The existing works either simply distill prior knowledge from the corresponding depth map to handle the RGB-image or blindly fuse color and geometric information to generate the depth-aware representations, hindering the performance of RGB-D saliency detectors. In this work, we introduce Cascade Graph Neural Networks (Cas-Gnn), a unified framework which is capable of comprehensively distilling and reasoning the mutual benefit between these two data sources through a set of cascade graphs, to learn powerful representations for RGB-D salient object detection. Cas-Gnn processes the two data sources separately and employs a novel Cascade Graph Reasoning (CGR) module to learn the powerful dense feature embeddings so that the saliency map can be easily inferred. Different from previous approaches, Cas-Gnn, by explicitly modeling and reasoning high-level relations between complementary data sources, can overcome many challenges like occlusions and ambiguities. Extensive experiments on several widely-used benchmarks demonstrate that CasGnn achieves significantly better performance than all existing RGB-D SOD approaches.