LapCR-Net: A Lightweight Monocular Depth Estimation Network via Laplacian Residual Reconstruction

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Monocular depth estimation is vital for autonomous driving and 3D reconstruction.
• Lightweight methods struggle with limited resources, shallow features, and insufficient cross-scale modeling, hindering structural consistency and detail preservation.

Purpose of the Study:

• To propose LapCR-Net, a lightweight network for monocular depth estimation.
• To address limitations in modeling cross-scale residual information for improved structural consistency and fine-grained detail recovery.

Main Methods:

• Developed a progressive Laplacian residual framework for coarse-to-fine multi-scale depth prediction.
• Introduced Structure-aware Feature Recalibration (SFR) and Depth-guided Convolution Module (DCM) for enhanced decoder feature representation.
• Implemented an uncertainty-driven collaborative refinement strategy for adaptive residual correction and artifact suppression.

Main Results:

• LapCR-Net achieves competitive performance with only 5.4 M parameters on NYU-Depth V2 and KITTI benchmarks.
• Demonstrated superior structural preservation and detail reconstruction capabilities.
• Achieved a favorable trade-off between accuracy and computational efficiency.

Conclusions:

• LapCR-Net effectively addresses the limitations of existing lightweight monocular depth estimation methods.
• The proposed network offers significant advantages in preserving scene structure and reconstructing fine details.
• LapCR-Net presents an efficient and accurate solution for real-world applications requiring precise depth information.