Cross360: 360° Monocular Depth Estimation via Cross Projections Across Scales.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Reconstruction

Background:

• 360° depth estimation is complex due to spherical image distortions and challenges in maintaining global continuity.
• Existing methods struggle to balance local and global consistency, with limited global perception in local features and discrepancies at patch boundaries.

Purpose of the Study:

• To propose Cross360, a novel architecture for accurate and globally consistent 360° depth estimation.
• To effectively integrate local and global information using less-distorted tangent patches and equirectangular features.

Main Methods:

• Developed a cross-attention-based architecture, Cross360, integrating local tangent patches and equirectangular features.
• Introduced a Cross Projection Feature Alignment module using cross-attention to align local features with global context.
• Implemented a Progressive Feature Aggregation with Attention module for refining multi-scaled features.

Main Results:

• Cross360 significantly outperforms existing methods on most benchmark datasets, particularly those with full 360° images.
• Demonstrated improved accuracy and global consistency in 360° depth estimation.
• Achieved state-of-the-art performance in challenging 360° depth estimation tasks.

Conclusions:

• Cross360 offers an effective solution for accurate and globally consistent 360° depth estimation.
• The proposed architecture successfully addresses limitations of previous methods in handling spherical image data.
• The method shows strong potential for applications requiring precise 3D scene understanding from panoramic imagery.