MSA-Net: Establishing Reliable Correspondences by Multiscale Attention Network.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Current deep learning methods for feature matching lack robustness due to outliers and insufficient information learning.
• This limits their effectiveness across diverse real-world scenarios.

Purpose of the Study:

• To develop a novel network, MSA-Net, for improved feature matching.
• Enhance robustness against outliers and improve representational ability of feature maps.
• Achieve effective inlier probability inference with reduced parameter count.

Main Methods:

• Proposed a multi-scale attention block to enhance outlier robustness.
• Designed context channel and spatial refine blocks for efficient information mining.
• Developed MSA-Net (Multi-Scale Attention Network) for feature matching.

Main Results:

• MSA-Net demonstrates superior performance in outlier removal and relative pose estimation.
• Achieved significant improvements over state-of-the-art methods on both indoor and outdoor datasets.
• Showcased an 11.7% improvement in relative pose estimation without RANSAC on the YFCC100M dataset.

Conclusions:

• MSA-Net offers enhanced robustness and effectiveness for feature matching tasks.
• The network achieves state-of-the-art performance with a reduced parameter count.
• Presents a promising solution for challenging computer vision applications requiring accurate feature correspondence.