C-NeRF: Representing Scene Changes as Directional Consistency Difference-Based NeRF.

Area of Science:

• Computer Vision
• 3D Scene Representation
• Artificial Intelligence

Background:

• Existing neural radiance fields (NeRFs) and 2D change detection (CD) methods struggle to accurately detect scene changes due to limitations in handling stereo information and appearance-based analysis.
• 2D CD methods often lead to false or missing detections by relying solely on pixel appearance differences between aligned image pairs, neglecting the volumetric nature of NeRFs.

Purpose of the Study:

• To develop a novel method for detecting object variations and scene changes within neural radiance fields (NeRFs).
• To address the limitations of current NeRF and 2D CD techniques in accurately identifying scene alterations.
• To enable precise scene monitoring and measurement applications by accurately predicting scene changes from arbitrary views.

Main Methods:

• Proposed C-NeRF (Change-NeRF), a method representing scene changes via directional consistency differences.
• Constructed two aligned NeRFs: one for the pre-change scene and one for the post-change scene.
• Implemented a change detection mechanism based on the constraint that real change points exhibit consistent representations across different view directions, unlike false positives.

Main Results:

• Developed a new dataset with ten diverse scenes to evaluate the proposed method.
• C-NeRF successfully generates change maps for arbitrarily specified view directions.
• The proposed approach significantly outperforms state-of-the-art 2D CD and NeRF-based methods in detecting scene changes.

Conclusions:

• C-NeRF effectively detects scene changes and object variations by leveraging directional consistency within NeRFs.
• The method overcomes the limitations of existing approaches, offering improved accuracy for scene monitoring and measurement.
• The developed dataset and C-NeRF approach provide a strong foundation for future research in dynamic scene understanding using NeRFs.