Semantic Correspondence: Unified Benchmarking and a Strong Baseline

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Semantic correspondence, matching keypoints across images with identical semantic meaning, is a complex computer vision challenge.
• Deep learning advancements have spurred progress, yet a comprehensive review of existing methods is lacking.

Purpose of the Study:

• To provide the first extensive survey and analysis of semantic correspondence methods.
• To introduce a novel taxonomy for classifying existing approaches.
• To establish a strong baseline for future research.

Main Methods:

• A new taxonomy is proposed to classify semantic correspondence methods based on their design.
• Existing methods are categorized and analyzed in detail.
• Comparative tables of benchmark results and controlled experiments are presented.

Main Results:

• The survey categorizes and analyzes various semantic correspondence techniques.
• Controlled experiments reveal the effectiveness of different method components.
• A new baseline achieves state-of-the-art performance on multiple benchmarks.

Conclusions:

• This survey offers a comprehensive reference for semantic correspondence research.
• The proposed baseline provides a solid foundation for future advancements in the field.
• Publicly available code facilitates reproducibility and further development.