Monocular Depth Estimation: Lightweight Convolutional and Matrix Capsule Feature-Fusion Network.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Current monocular depth estimation algorithms using Convolutional Neural Networks (CNNs) exhibit limited adaptability to angle transformations, impacting accuracy and robustness.
• Existing methods often lack the precision and reliability needed for diverse real-world applications.

Purpose of the Study:

• To develop a lightweight network, CNNapsule, that enhances adaptability to angle transformations in monocular depth estimation.
• To improve the accuracy and robustness of depth estimation compared to existing CNN-based approaches.

Main Methods:

• Proposed a novel network architecture, CNNapsule, integrating Convolutional Neural Network (CNN) features with matrix capsule features.
• Introduced a fusion block module for improved adaptability to perspective transformations.
• Developed a custom loss function considering dataset characteristics like long-tail distribution, gradient similarity, and structural similarity.

Main Results:

• Achieved superior accuracy on C1 and C2 indices and better visual results on NYU Depth V2 and KITTI datasets compared to traditional and non-transfer learning deep learning methods.
• Demonstrated a 65% reduction in trainable parameters compared to existing literature methods.
• Validated generalization capabilities using internet-sourced and mobile phone data.

Conclusions:

• CNNapsule offers a more accurate, robust, and computationally efficient solution for monocular depth estimation, particularly in handling angle variations.
• The proposed method shows significant improvements in performance and parameter efficiency, making it suitable for real-world applications.
• The network's generalization ability is confirmed across diverse datasets, highlighting its practical utility.