PGFNet: Preference-Guided Filtering Network for Two-View Correspondence Learning.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Feature matching in computer vision relies on accurate image correspondences.
• Existing methods struggle with outliers in initial correspondences, hindering contextual information capture.

Purpose of the Study:

• To propose a novel network, Preference-Guided Filtering Network (PGFNet), for robust correspondence selection and camera pose estimation.
• To address the challenge of outliers in feature matching for improved vision tasks.

Main Methods:

• Developed an iterative filtering structure to learn correspondence preference scores for outlier mitigation.
• Introduced a Grouped Residual Attention block with feature grouping and attention mechanisms to enhance score reliability.
• Employed ablation studies and comparative experiments for evaluation.

Main Results:

• PGFNet effectively selects correct correspondences and recovers accurate camera poses.
• The proposed methods significantly outperform state-of-the-art techniques on challenging datasets.
• Demonstrated performance gains in both outlier removal and camera pose estimation tasks.

Conclusions:

• PGFNet offers a superior approach to correspondence selection and camera pose estimation.
• The iterative filtering and attention mechanisms are key to handling outliers and improving reliability.
• The method shows strong potential for various feature matching-based computer vision applications.