What is optimized in convex relaxations for multilabel problems: connecting discrete and continuously inspired MAP...

Area of Science:

• Computer Vision
• Image Processing
• Optimization

Background:

• Markov random fields (MRFs) are widely used for image labeling tasks.
• Existing MRF relaxations can be biased by grid geometry.
• Continuous convex relaxations offer an alternative with reduced bias.

Purpose of the Study:

• To unify Markov random fields (MRFs) with continuous tight convex relaxations.
• To enhance understanding of discrete relaxations through continuous methods.
• To develop efficient optimization techniques for multilabel image assignment.

Main Methods:

• Unified MRFs and continuous tight convex relaxations.
• Developed two novel minimization schemes for efficient optimization.
• Generalized energy formulation to nonmetric and orientation-dependent smoothness terms.

Main Results:

• Continuous methods are nonlinear extensions of MRF relaxations, offering less grid bias.
• Proposed methods improve minimization efficiency for tight formulations.
• Demonstrated utility of new schemes through numerical experiments.

Conclusions:

• The unified framework provides deeper insights into discrete relaxations.
• New optimization strategies enhance the applicability of continuous methods.
• Generalization of energy terms broadens the scope of MRF-based image analysis.