Enhancing long-range depth estimation via heterogeneous CNN-transformer encoding and cross-dimensional semantic fusion

Abstract

Abstract Monocular depth estimation enables 3D scene reconstruction from a single 2D image, offering a cost-effective solution widely applied in autonomous driving and UAVs. However, existing deep neural networks often fail to balance local texture details with global contextual information, leading to significant inaccuracies in distant-region depth prediction. To address this challenge, we introduce a novel monocular depth estimation framework featuring a heterogeneous encoder and a Cross-dimensional Semantic Fusion (CSF) module. The heterogeneous encoder integrates the initial convolutional layers of ResNet-50 with the hierarchical attention mechanism of Swin Transformer to efficiently capture both local details and long-range dependencies. Specifically targeting the characteristics of distant objects—low pixel occupancy but high semantic relevance—the CSF module enhances feature aggregation in the decoder through multi-scale interactions and spatial-channel coupling. Additionally, the decoder incorporates a Depth-Separable Upsampling Block (DSUB) and a Multi-scale Self-Attention (MSA) module to refine detail restoration and ensure spatial consistency. Experiments validate the superiority of our method. On the KITTI dataset, it achieves leading results: 0.050 Abs-Rel, 2.107 RMSE, and a long-range error of 0.2725. The SUN RGB-D dataset demonstrates strong generalization with an Abs-Rel of 0.142. This framework significantly advances long-range depth estimation research and shows broad application prospects.