Seed to Prune: A Seeded Graph Neural Network for Two-View Correspondence Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Geometric Deep Learning

Background:

• Correspondence learning is crucial for tasks like image matching and 3D reconstruction.
• Existing methods often struggle with outliers and may be biased by global or local context extraction.
• There is a need for robust methods that can effectively handle noisy correspondence data.

Purpose of the Study:

• To propose a novel and effective method for correspondence learning named SGNNet.
• To address the limitations of existing methods in handling outliers and biased contextual information.
• To achieve state-of-the-art performance in correspondence learning across various datasets.

Main Methods:

• SGNNet employs a dynamic seeding module to sample reliable matches as seeds.
• An intraseed attention module (ISAM) captures geometrical relations among seeds to enhance their features.
• A dynamic unseeding module aggregates contextual information from seeds and broadcasts it to original matches.

Main Results:

• SGNNet effectively rejects outliers from putative correspondences.
• The method achieves new state-of-the-art (SOTA) scores across multiple domains and datasets.
• On YFCC100M, SGNNet attained an AUC@5° of 56.43% without RANSAC, surpassing prior work by 4.51 absolute percentage points.

Conclusions:

• SGNNet offers a simple yet highly effective approach to correspondence learning.
• The proposed method demonstrates superior performance in outlier rejection and accuracy.
• SGNNet sets a new benchmark for correspondence learning, particularly in large-scale image datasets.