A Review of Depth and Normal Fusion Algorithms

Area of Science:

• Computer Vision
• Geometric Reconstruction
• 3D Data Processing

Background:

• Depth maps and surface normals are crucial for 3D scene understanding.
• Existing methods for combining these geometric cues have limitations in accuracy and robustness.
• Variations in surface normal formulations impact depth reconstruction performance.

Purpose of the Study:

• To systematically review and analyze algorithms for combining depth and surface normal information.
• To introduce novel methods for enhanced depth map refinement.
• To evaluate the performance of new methods against existing techniques.

Main Methods:

• Comparative analysis of depth reconstruction algorithms using surface normals.
• Development of a generalized least squares fusion method with novel normal weighting.
• Implementation of a novel method based on Total Generalized Variation (TGV).

Main Results:

• The generalized fusion method, with its novel weighting, outperforms previous approaches in the depth error domain.
• The TGV-based method achieves superior performance in geodesic normal distance error.
• Both new methods demonstrate significant improvements in geometric surface reconstruction accuracy.

Conclusions:

• Combining depth and surface normal information is key to refining depth maps.
• The proposed generalized fusion and TGV methods offer state-of-the-art performance.
• Accurate surface normal representation is critical for effective depth reconstruction.