Abstract
Self-supervised monocular depth estimation has gained significant traction in autonomous driving and robotics owing to its inherent advantages. Recent studies focus on joint semantic and depth information utilization to improve depth estimation precision. This cross-modal fusion represents a promising approach to enhance depth estimation quality. However, achieving effective fusion and interaction between semantic and depth features remains a significant challenge. We propose a novel self-supervised framework integrating cross-modal feature interaction strategy with edge optimization strategy to address this. The proposed architecture comprises two key components: a pose estimation network and a feature interaction network. Within the feature interaction network, we introduce a cross-modal interaction module that enables bidirectional refinement of depth and semantic features; this module facilitates enhanced synergistic integration of multimodal information. Furthermore, we design a semantic-guided edge optimization module that harnesses semantic boundary features to augment depth representations, producing refined edge details. Extensive experiments conducted on the KITTI dataset demonstrate that the proposed model achieves competitive performance in monocular depth estimation, and compared to other state-of-the-art research methods, our approach exhibits superior performance.
Funding Information
  • Science and Technology Program Project of Tianjin (No. 21YFFCYS00080)

This publication has 21 references indexed in Scilit: