Point Light Source Position Estimation From RGB-D Images by Learning Surface Attributes.

Area of Science:

• Computer Vision
• Photometry
• Robotics

Background:

• Estimating light source position (LSP) is crucial in computer vision but challenging due to varying surface properties.
• Traditional methods often assume Lambert's law, which is inaccurate for many real-world surfaces.
• Existing techniques may not effectively utilize geometric and photometric surface information for LSP estimation.

Purpose of the Study:

• To develop a more robust LSP estimation method that accounts for diverse surface properties.
• To improve the accuracy of LSP estimation in RGB-D video sequences by incorporating camera pose.
• To outperform current state-of-the-art methods in LSP estimation accuracy.

Main Methods:

• Classifying image surface segments based on photometric and geometric attributes (e.g., glossy, matte, curved).
• Assigning weights to surface segments according to their suitability for LSP estimation.
• Utilizing estimated camera pose for global constraint of LSP in RGB-D video sequences.

Main Results:

• The proposed method significantly outperforms state-of-the-art techniques on benchmark and custom RGB-D datasets.
• Surface weighting based on attributes reduces angular error from 12.6° to 8.2° (Boom dataset) and 24.6° to 4.8° (RGB-D video dataset).
• Global constraint using camera pose further enhances accuracy, achieving 4.8° error compared to 8.5° with single frames.

Conclusions:

• Classifying and weighting image surface segments by their attributes is superior to assuming uniform surface properties for LSP estimation.
• Integrating camera pose for global constraint in RGB-D videos substantially improves LSP estimation accuracy.
• The developed approach offers a more accurate and reliable solution for light source position estimation in complex visual scenes.