TopicFM+: Boosting Accuracy and Efficiency of Topic-Assisted Feature Matching

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Image matching is challenging in scenes with significant variations or limited texture.
• Existing Transformer-based methods for global scene context encoding are computationally expensive and may miss high-level contextual information like spatial structures or semantic shapes.

Purpose of the Study:

• To develop a computationally efficient image matching method that effectively captures high-level contextual information.
• To overcome the limitations of existing Transformer-based approaches in terms of computational cost and contextual understanding.

Main Methods:

• A novel image matching method leveraging a topic-modeling strategy to represent images as distributions over semantic topics.
• A coarse-level matching network using attention on fixed-sized topics and small features for efficiency.
• A dynamic feature refinement network for precise fine-level matching.

Main Results:

• The proposed method demonstrates superior performance in challenging image matching scenarios.
• Achieved top 9% ranking in the Image Matching Challenge 2023 without ensemble techniques.
• Reduced computational costs by approximately 50% compared to Transformer-based methods.

Conclusions:

• The topic-modeling approach effectively captures comprehensive context and generates discriminative features for image matching.
• The method offers a significant improvement in computational efficiency while maintaining high accuracy in difficult scenarios.
• This work provides a promising direction for efficient and effective image matching.