Online learning of correspondences between images

Area of Science:

• Computer Vision
• Machine Learning
• Geometry

Background:

• Establishing point correspondences between images is crucial for 3D reconstruction and scene understanding.
• Existing methods often rely on strong assumptions about imaging geometry or require computationally expensive global optimization.
• A need exists for robust methods that handle general imaging conditions and unknown scene geometry.

Purpose of the Study:

• To develop a novel iterative learning method for estimating point correspondences between image sequences.
• To address the challenge of unknown projection geometry and surface shape.
• To achieve real-time performance with high accuracy and fast convergence.

Main Methods:

• Iterative learning of point correspondences using sequences of point-set pairs.
• Optimization based on Neyman's chi-square divergence between uncertainty-representing densities.
• Representation of densities using channel vectors computed via a basis function approach.
• Updating the mapping between vectors with each new image pair.

Main Results:

• The proposed method achieves fast convergence and high accuracy.
• The algorithm operates in real time.
• Experimental results demonstrate superiority over state-of-the-art methods in convergence and accuracy.

Conclusions:

• The novel iterative method effectively learns point correspondences for image sequences with general imaging geometry.
• The approach offers a robust and efficient solution for real-time applications.
• This work advances the state-of-the-art in correspondence estimation for uncalibrated vision systems.