Convolutional Neural Network Architecture for Geometric Matching.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Geometric matching is crucial for image analysis, involving establishing correspondences aligned with geometric models like affine or homography transformations.
• Estimating transformation parameters and identifying inliers are key challenges in traditional geometric matching pipelines.

Purpose of the Study:

• To develop an end-to-end trainable convolutional neural network architecture for robust geometric matching.
• To enable accurate estimation of geometric model parameters and simultaneous inlier detection.
• To improve generalization capabilities for matching unseen images.

Main Methods:

• A novel convolutional neural network architecture is proposed, integrating feature extraction, matching, and parameter estimation.
• The network is trained end-to-end using synthetically generated imagery, eliminating the need for manual annotations.
• A specialized matching layer is incorporated to enhance generalization.

Main Results:

• The proposed network achieves state-of-the-art performance on challenging datasets including PF, TSS, and Caltech-101.
• The model demonstrates strong generalization capabilities to novel images due to the proposed matching layer.
• The same model effectively performs both instance-level and category-level matching.

Conclusions:

• The developed convolutional neural network offers an effective and efficient solution for geometric matching and parameter estimation.
• Training with synthetic data and employing an advanced matching layer are key to achieving high generalization.
• The model's versatility in handling both instance and category matching advances the field of computer vision.