MogaDepth: Multi-Order Feature Hierarchy Fusion for Lightweight Monocular Depth Estimation

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Monocular depth estimation is vital for autonomous driving and augmented reality.
• Existing lightweight methods often overlook crucial mid-order semantic feature interactions.

Purpose of the Study:

• To introduce MogaDepth, a novel lightweight architecture for improved monocular depth estimation.
• To enhance the representation of mid-level features for greater depth accuracy.

Main Methods:

• Developed the Continuous Multi-Order Gated Aggregation (CMOGA) module to enhance mid-level features.
• Introduced MambaSync for efficient global-local feature communication.
• Proposed MogaDepth, a lightweight and expressive network architecture.

Main Results:

• MogaDepth achieved competitive or superior performance on the KITTI benchmark, improving error metrics.
• Outperformed existing methods on the Make3D benchmark, demonstrating robustness to domain shifts and challenging scenarios.
• Achieved up to 13% faster inference on edge devices without performance compromise.

Conclusions:

• MogaDepth offers an effective and efficient solution for real-world monocular depth estimation.
• The proposed CMOGA and MambaSync modules significantly contribute to improved depth accuracy and efficiency.