CNN–Transformer Fusion Network With Guided Deconvolution Block for Edge-Aware Depth Completion

Depth completion has become a fundamental technology in computer vision and autonomous driving systems, enabling more accurate and detailed 3D mapping by reconstructing fine geometric structures, improving depth accuracy, and enhancing spatial continuity for applications such as autonomous navigation and environmental understanding. In particular, sensors like LiDAR inherently generate sparse depth measurements, making effective reconstruction crucial for achieving comprehensive and high-fidelity 3D representations in both indoor and outdoor environments. This process aims to recover dense maps from sparse data while preserving geometric details and structural consistency. Most existing methods use convolutional networks, which capture local feature correlations effectively but often struggle with longrange dependencies between non-adjacent pixels. To overcome the limitation of convolutional networks in capturing long-range dependencies and global context, we propose a novel architecture that integrates a Transformer with a convolutional encoder, thereby combining global attention mechanisms with local inductive biases. The decoder employs a CNN augmented with a Guided Deconvolution Block (GDB) embedding the GDB into its upsampling stages enables effective recovery of spatial details and preservation of sharp edges, which are typically degraded during the encoding process. By fusing features from the CNN, which captures local structural information, with the Transformer, which encodes global contextual relationships, this architectural fusion effectively preserves spatial features across different scales and maintains clear object boundaries, leading to more accurate depth reconstruction. We thoroughly evaluate the proposed network on both standard benchmarks: the indoor NYU Depth v2 dataset and the outdoor KITTI Depth Completion dataset. On NYU Depth v2, the model demonstrates strong performance, approaching state-of-the-art results, while effectively capturing fine spatial details. On KITTI, the model achieves robust predictions in sparse and long-range regions, accurately estimating depth values even for distant or sparsely measured points and maintaining consistency across object boundaries. These results illustrate that the proposed approach generalizes effectively across diverse environments, from structured indoor spaces to complex, dynamic outdoor driving scenarios, demonstrating robustness to sparsity, noise, and variations in scene geometry. Overall, the combination of Transformer-based global reasoning, CNN-based local feature extraction, and GDB-enhanced decoding provides a powerful and efficient solution for dense 3D mapping across diverse environments.