A novel factor graph-based optimization technique for stereo correspondence estimation

Area of Science:

• Computer Vision
• Artificial Intelligence
• 3D Reconstruction

Background:

• Accurate dense disparity estimation is crucial for 3D scene reconstruction from stereo images.
• Challenges include homogeneous textures, varying illumination, and occlusions, which limit traditional methods like Markov random fields.
• Learning-based methods offer rich features but can be complex.

Purpose of the Study:

• To develop a new probabilistic graphical model for improved disparity estimation.
• To address limitations of fixed neighborhood systems in previous methods.
• To enhance accuracy in challenging stereo image conditions.

Main Methods:

• A novel factor graph-based probabilistic graphical model was developed.
• The model incorporates a spatially variable neighborhood structure adapting to local scene characteristics.
• Spatial dependencies among scene characteristics and disparity estimates are enforced.

Main Results:

• The proposed factor graph algorithm achieved higher accuracy in disparity estimation compared to state-of-the-art non-learning and learning-based methods.
• Performance was validated on Middlebury benchmark stereo datasets.
• The method demonstrated improved accuracy with varying texture and illumination.

Conclusions:

• The factor graph model offers a flexible and accurate approach to dense disparity estimation.
• It effectively handles complex dependency structures and improves computational efficiency.
• The framework is adaptable for other dense estimation tasks like optical flow.