A simple monocular depth estimation network for balancing complexity and accuracy.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Monocular depth estimation is vital for visual tasks but often requires high computational cost.
• Existing methods struggle to balance accuracy with efficiency in practical applications.

Purpose of the Study:

• To develop a computationally efficient and accurate monocular depth estimation model.
• To address the trade-off between performance and resource consumption in depth prediction.

Main Methods:

• Proposing SimMDE, a novel model treating monocular depth estimation as ordinal regression.
• Utilizing a Deformable Cross-Attention Feature Fusion (DCF) decoder with sparse attention to reduce Transformer complexity.
• Introducing Local Multi-dimensional Convolutional Attention (LMC) and Wavelet Attention Transformer (WAT) modules for enhanced feature extraction and pixel-level classification.

Main Results:

• SimMDE achieves exceptional accuracy on NYU and KITTI benchmark datasets.
• The model demonstrates significant improvements in absolute relative error (AbsRel) by 11.7% (NYU) and 10.3% (KITTI).
• SimMDE achieves these results with fewer parameters and high computational efficiency.

Conclusions:

• SimMDE offers a promising solution for accurate and efficient monocular depth estimation.
• The proposed model effectively integrates multi-scale features and enhances local feature extraction.
• SimMDE represents a significant advancement over existing methods, particularly for real-world applications.