Density-Guided Incremental Dominant Instance Exploration for Two-View Geometric Model Fitting.

Area of Science:

• Computer Vision
• Geometric Modeling
• Machine Learning

Background:

• Existing two-view multi-model fitting methods use disjointed model generation and selection steps.
• This two-step approach leads to redundant computations and lacks efficiency.

Purpose of the Study:

• To develop a real-time algorithm for accurate geometric model fitting and data segmentation.
• To address the limitations of traditional two-step fitting methods by integrating generation and selection.

Main Methods:

• Introduced D2Fitting, an algorithm that iteratively explores dominant instances by alternating model generation and selection.
• Developed a density-guided sampler for high-quality minimal subset sampling.
• Implemented a global-residual optimization for refining sampled subsets and mitigating noise.

Main Results:

• D2Fitting achieves real-time performance by avoiding redundant instance generation.
• The algorithm accurately estimates the number and parameters of geometric models.
• Simultaneous and efficient segmentation of input data is achieved.

Conclusions:

• D2Fitting significantly outperforms existing state-of-the-art methods in geometric model fitting and data segmentation.
• The integrated approach of model generation and selection offers substantial computational and accuracy improvements.