Unsupervised Monocular Depth Estimation Method Based on Multimodal Interaction Perception

Unsupervised monocular depth estimation has become a mainstream approach in depth estimation due to its ability to eliminate the reliance on expensive ground truth depth annotations, making it more scalable for real-world applications. However, despite significant progress, two major challenges remain: the difficulty in effectively capturing both global context and fine-grained local details, and the limited generalization ability in complex, dynamic scenes.To address these issues, we propose MIPDepth, an unsupervised monocular depth estimation framework based on multi-modal interaction perception. Our method leverages the complementary strengths of convolutional neural networks (CNNs) and vision transformers (ViTs) to overcome their individual limitations. Specifically, the Multi-Scale Pyramid Interaction Perception Layer (MPIPL) is designed to extract rich multi-scale features, while the Multi-Modal Interaction Perception Module (MIPM) enables effective interaction between CNN and ViT streams, enhancing both local detail preservation and global contextual understanding.Extensive experiments on KITTI and Cityscapes benchmarks demonstrate that MIPDepth achieves state-of-the-art performance, especially in dynamic environments, validating its superiority in accuracy, robustness, and generalization.