Discrete-Continuous Transformation Matching for Dense Semantic Correspondence.

Area of Science:

• Computer Vision
• Machine Learning
• Image Analysis

Background:

• Existing dense semantic correspondence techniques struggle with complex geometric variations, particularly affine transformations.
• The large solution space of affine transformations presents a significant computational challenge for current methods.

Purpose of the Study:

• To develop a robust framework for dense semantic correspondence that accurately addresses complex geometric deformations.
• To enable efficient computation of dense affine transformation fields for improved image matching accuracy.

Main Methods:

• Introduced a discrete-continuous transformation matching (DCTM) framework.
• Inferred dense affine transformation fields via discrete label optimization with continuous regularization.
• Utilized constant-time edge-aware filtering and an affine-varying CNN-based descriptor for efficient computation.
• Incorporated correspondence consistency and confidence-guided filtering to enhance convergence.

Main Results:

• The proposed DCTM framework demonstrates superior performance compared to state-of-the-art methods for dense semantic correspondence.
• The method effectively handles complex geometric variations, including affine transformations, in semantically similar images.
• Experimental results validate the model's efficacy across various benchmarks and applications.

Conclusions:

• The DCTM framework offers a practical and efficient solution for dense semantic correspondence with complex geometric variations.
• The approach significantly advances the capabilities of image matching and analysis in the presence of deformations.