Graph-Based Depth Denoising & Dequantization for Point Cloud Enhancement.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Signal Processing

Background:

• 3D point clouds are crucial for various applications but are often degraded by sensor noise and quantization.
• Existing methods denoise point clouds after 3D reconstruction, potentially missing opportunities for early error correction.
• Depth measurement imperfections arise from signal-dependent noise and non-uniform quantization during sensing.

Purpose of the Study:

• To develop a novel method for enhancing depth measurements directly from sensor images prior to 3D point cloud synthesis.
• To improve the quality of 3D point clouds by addressing noise and quantization at the image acquisition stage.
• To introduce a depth formation model that accounts for sensor-specific noise and quantization characteristics.

Main Methods:

• Modeled depth formation as signal-dependent noise and log-based quantization.
• Developed a graph-based approach to encode intra-view and inter-view similarities in depth images.
• Formulated a convex and differentiable maximum a posteriori (MAP) graph filtering objective.
• Utilized accelerated gradient descent (AGD) with Gershgorin Circle Theorem (GCT) for efficient optimization.

Main Results:

• The proposed depth enhancement method significantly improved 3D point cloud quality.
• Experimental results demonstrated superior performance compared to state-of-the-art point cloud and image denoising techniques.
• The depth formation model was validated using empirical data from a representative depth sensor.

Conclusions:

• Enhancing depth measurements a priori, close to the physical sensing process, is more effective than post-hoc denoising.
• The developed MAP graph filtering approach effectively reduces noise and quantization artifacts in depth images.
• This work offers a promising direction for improving the fidelity of 3D point cloud reconstruction from noisy depth data.