Temporal and Spatial Denoising of Depth Maps.

Area of Science:

• Computer Vision
• Image Processing
• 3D Sensing

Background:

• RGB-D cameras provide depth data but suffer from occlusions and inaccuracies.
• High-resolution depth map acquisition is common, yet data quality remains a challenge.
• Existing methods struggle with refining noisy or incomplete depth information.

Purpose of the Study:

• To propose a novel method for refining depth maps acquired by RGB-D cameras.
• To address issues like occlusion, inaccurate depth values, and temporal variations.
• To enhance the image quality of depth data for improved 3D reconstruction and analysis.

Main Methods:

• An exemplar-based inpainting technique is adapted and modified for depth map refinement.
• The method leverages concepts from image inpainting to fill occluded or erroneous depth regions.
• The approach is evaluated using the Tsukuba Stereo Dataset and self-recorded data.

Main Results:

• The proposed method successfully removes artifacts from RGB-D depth maps.
• Experimental results demonstrate significant improvements in depth data quality.
• Evaluation metrics include peak signal-to-noise ratio and computational time.

Conclusions:

• The enhanced exemplar-based inpainting method is effective for refining RGB-D depth maps.
• This technique offers a viable solution for improving the accuracy and reliability of depth sensing.
• The refined depth data has potential applications in various 3D imaging fields.