MonoPrior-Fusion: Monocular-Prior-Guided Multi-Frame Depth Estimation with Multi-Scale Geometric Fusion.

Area of Science:

• Computer Vision
• Robotics
• 3D Perception

Background:

• Accurate 3D perception is vital for indoor robotics, augmented reality, and autonomous navigation.
• Existing multi-frame depth estimation methods struggle with challenging indoor conditions like weak textures and complex layouts.

Purpose of the Study:

• To develop a novel framework, MonoPrior-Fusion (MPF), to enhance multi-frame depth estimation in indoor environments.
• To address performance degradation caused by challenging indoor scene characteristics.

Main Methods:

• MPF integrates pixel-wise monocular priors into the multi-view matching process.
• A hierarchical fusion architecture and a geometric consistency loss based on virtual planes are employed.
• Cost-volume hypotheses are modulated to disambiguate matches.

Main Results:

• MPF significantly outperforms state-of-the-art multi-frame baselines on benchmark datasets (ScanNetV2, 7Scenes, TUM RGB-D, GMU Kitchens).
• The method demonstrates strong generalization capabilities across unseen indoor domains.
• Integration into volumetric fusion pipelines yields more accurate and complete 3D reconstructions.

Conclusions:

• MonoPrior-Fusion (MPF) effectively enhances multi-frame depth estimation for indoor robotics and AR applications.
• The framework provides robust and accurate 3D perception, outperforming existing methods in challenging scenarios.
• MPF shows significant promise for dense 3D reconstruction and mapping tasks.