Convolutional Hough Matching Networks for Robust and Efficient Visual Correspondence.

Area of Science:

• Computer Vision
• Machine Learning
• Geometric Deep Learning

Background:

• Establishing reliable visual correspondences is challenging due to large image variations.
• Geometric relations are crucial for robust image matching.
• Existing feature representation methods often struggle with significant intra-class variations.

Purpose of the Study:

• To propose an effective geometric matching algorithm using a Hough transform perspective on convolutional matching.
• To develop a trainable neural layer for non-rigid matching with interpretable parameters.
• To improve the efficiency of high-dimensional voting in geometric matching.

Main Methods:

• Introduced Convolutional Hough Matching (CHM) by distributing similarities over a geometric transformation space.
• Developed a trainable neural layer with a semi-isotropic high-dimensional kernel for non-rigid matching.
• Proposed an efficient kernel decomposition with center-pivot neighbors to sparsify kernels without performance loss.

Main Results:

• Achieved state-of-the-art performance on standard benchmarks for semantic visual correspondence.
• Demonstrated strong robustness to challenging intra-class variations in images.
• Validated the effectiveness of CHM layers in a neural network for translation and scaling matching.

Conclusions:

• Convolutional Hough Matching (CHM) offers a powerful approach for robust visual correspondence.
• The proposed method effectively leverages geometric relations for improved matching accuracy.
• CHM represents a significant advancement in handling large image variations and intra-class differences.