Multi-view scene matching with relation aware feature perception.

Area of Science:

• Computer Vision
• Machine Learning
• Geospatial Analysis

Background:

• Scene matching with multi-source and multi-view data presents significant challenges in metric feature extraction.
• Existing methods struggle with coherence and metric consistency across diverse datasets.

Purpose of the Study:

• To propose a Siamese network model for robust Spatial Relation Aware feature perception and fusion.
• To enhance the reliability and accuracy of metric models in complex scene matching scenarios.

Main Methods:

• Developed a relation-aware feature perception using spatial relation vector decomposition for horizontal and vertical consistency.
• Implemented a large-scale local information perception strategy for optimal scale selection in aerial and satellite imagery.
• Proposed a feature selection and fusion strategy, assessing backbone network levels for improved metric confidence.

Main Results:

• The proposed method demonstrates enhanced reliability in metric models for scene matching.
• Experimental results on the University-1652 dataset and real-world data validate the approach's efficacy.
• The model successfully improves the representation of pivotal components within metric features.

Conclusions:

• The Spatial Relation Aware Siamese network effectively addresses challenges in multi-source, multi-view scene matching.
• The method offers a reliable solution for enhancing metric feature extraction and fusion.
• This approach shows potential for broad applicability across various scene matching tasks.