Neural Colour Correction for Indoor 3D Reconstruction Using RGB-D Data

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• Generating photo-realistic 3D models is crucial for human-centered applications.
• Multiview 3D reconstruction of indoor scenes suffers from color inconsistencies due to varying acquisition conditions.
• These inconsistencies lead to visual artifacts in the final 3D models.

Purpose of the Study:

• To propose a novel neural-based approach for color correction in indoor 3D reconstruction.
• To harmonize colors from sparse captures in complex indoor environments.
• To address the challenge of generating photo-realistic 3D models.

Main Methods:

• A lightweight and efficient neural network approach is developed.
• A fully connected deep neural network learns an implicit representation of color in 3D space.
• Camera-dependent effects are captured, and transformations are estimated to regenerate pixels.

Main Results:

• The proposed method effectively harmonizes color from sparse captures.
• It outperforms existing state-of-the-art approaches on the MP3D dataset.
• The approach generates visually appealing and artifact-free 3D models.

Conclusions:

• The neural-based color correction method is effective for indoor 3D reconstruction.
• It offers a significant improvement over current methods for generating photo-realistic models.
• The approach is lightweight, efficient, and suitable for complex indoor scenes.