AGDF-Net: Learning Domain Generalizable Depth Features With Adaptive Guidance Fusion

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Cross-domain generalizable depth estimation aims to predict depth in real-world scenes using models trained on synthetic data.
• Existing methods struggle with domain gaps, limiting the extraction of crucial depth information and increasing interference.
• This hinders the performance of depth estimation models in real-world applications.

Purpose of the Study:

• To propose a novel domain generalizable feature extraction network with adaptive guidance fusion (AGDF-Net).
• To fully acquire essential features for depth estimation at multi-scale feature levels, overcoming limitations of previous methods.
• To achieve state-of-the-art performance in cross-domain depth estimation using only synthetic training data.

Main Methods:

• AGDF-Net separates images into initial depth and weak-related depth components using reconstruction and contrary losses.
• An adaptive guidance fusion module intensifies initial depth features for domain generalizable intensified depth features.
• The intensified features are then used by an arbitrary depth estimation network for real-world depth prediction.

Main Results:

• AGDF-Net achieves state-of-the-art performance on diverse real-world datasets (KITTI, NYUDv2, NuScenes, DrivingStereo, CityScapes) when trained solely on synthetic data.
• Experiments in a semi-supervised setting with limited real-world data further confirm AGDF-Net's superiority.
• The method effectively addresses the domain gap challenge in cross-domain depth estimation.

Conclusions:

• AGDF-Net offers a robust solution for cross-domain generalizable depth estimation.
• The proposed adaptive guidance fusion mechanism enhances feature extraction capabilities.
• This approach enables accurate real-world depth estimation using synthetic training data, paving the way for more efficient and versatile depth sensing applications.