ConvMatch: Rethinking Network Design for Two-View Correspondence Learning.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Multilayer perceptron (MLP) is a common backbone for two-view correspondence learning, excelling at individual feature extraction but lacking inherent context.
• Existing methods often append context-capturing modules to MLPs, yet performance remains limited by the backbone's native inability to aggregate context.

Purpose of the Study:

• To design a novel correspondence learning network, ConvMatch, utilizing a convolutional neural network (CNN) backbone for inherent context aggregation.
• To improve the accuracy of motion vector estimation by implicitly regularizing putative motion vectors through a dense motion field.

Main Methods:

• ConvMatch employs a CNN backbone, enabling inherent context aggregation for correspondence learning.
• It converts sparse motion vectors to a dense motion field, regularizes putative vectors, and uses CNNs to rectify local errors caused by outliers.
• Global information injection and bilateral convolution are introduced to better fit spatial transformations and handle motion field discontinuities.

Main Results:

• ConvMatch demonstrates superior performance in estimating correct motion vectors from rectified motion fields.
• The network consistently outperforms state-of-the-art methods in key computer vision tasks.

Conclusions:

• ConvMatch offers a significant advancement in correspondence learning by leveraging CNNs for effective context aggregation.
• The proposed methods achieve state-of-the-art results in relative pose estimation, homography estimation, and visual localization.