Saliency Driven Monocular Depth Estimation Based on Multi-scale Graph Convolutional Network

Monocular depth estimation is a fundamental and crucial task in computer vision that enables scene understanding from a single image. This paper proposes a novel approach for saliency-driven monocular depth estimation based on a multi-scale Graph Convolutional Network (GCN). Our method utilizes saliency information to guide the depth estimation process and employs a multi-scale GCN to capture local and global contextual cues. The proposed framework constructs a graph structure using RGB images to represent the relationships between image regions. We designed a multi-scale feature fusion module called DS Fusion, by applying GCN at multiple scales, our method effectively integrates depth features and saliency features to predict accurate depth maps. Extensive experiments conducted on KITTI and NYU datasets demonstrate the superior performance of our approach compared to state-of-the-art techniques. Additionally, we perform indepth analysis of the network architecture and discuss the impact of saliency cues on depth estimation accuracy. Our proposed method showcases the potential of combining saliency information and GCN in monocular depth estimation, contributing to the progress of scene understanding and depth perception from a single image.