Self-Supervised Depth Completion Guided by 3D Perception and Geometry Consistency.

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• Depth completion is vital for computer graphics and vision.
• Unsupervised high-precision depth completion remains challenging due to ignored 3D structure.
• Existing unsupervised methods suffer from inaccurate spatial propagation and mixed-depth issues.

Purpose of the Study:

• To develop a high-precision self-supervised depth completion method.
• To address limitations of previous unsupervised approaches by incorporating 3D structural information.
• To leverage multi-view geometry consistency for accurate depth map prediction.

Main Methods:

• Utilized 3D perceptual features and multi-view geometry consistency.
• Developed a 3D perceptual spatial propagation using point clouds and attention mechanisms.
• Introduced multi-view geometric constraints for self-supervised optimization.

Main Results:

• Achieved state-of-the-art depth completion performance on benchmark datasets (NYU-Depth-v2, VOID, KITTI).
• Outperformed existing unsupervised depth completion methods.
• Demonstrated competitive performance compared to supervised methods.

Conclusions:

• The proposed self-supervised method effectively addresses challenges in unsupervised depth completion.
• Incorporating 3D perceptual features and multi-view consistency enhances depth map accuracy.
• The method offers a promising alternative to supervised approaches for depth completion tasks.