Robust Model Fitting Using Higher Than Minimal Subset Sampling.

Area of Science:

• Computer Vision
• Machine Learning
• Data Science

Background:

• Identifying underlying models in noisy, outlier-contaminated data is crucial in computer vision.
• Complex cost functions often necessitate approximate solutions from discrete parameter spaces.
• Minimal subset sampling can yield hypotheses distant from the true model due to noise.

Purpose of the Study:

• To investigate the effectiveness of using higher than minimal subset sampling for hypothesis generation.
• To develop a computationally tractable method for robust model fitting using larger subsets.
• To improve the accuracy and efficiency of identifying underlying models in challenging datasets.

Main Methods:

• Empirical studies on increasing sample size beyond minimal (p) to p+2 for inlier subsets.
• Development of a novel robust model fitting method starting from arbitrary hypotheses.
• Inclusion of a stopping criterion for adequate accuracy to save computational time.

Main Results:

• Increasing sample size to p+2 significantly improves the probability of generating hypotheses closer to the true model from inliers.
• The probability of selecting all-inlier samples decreases rapidly with increased sample size.
• The proposed method demonstrates accuracy and efficiency compared to state-of-the-art techniques on synthetic and real data.

Conclusions:

• Higher than minimal subset sampling, particularly up to p+2 from inliers, is feasible and beneficial for hypothesis generation.
• The proposed robust model fitting method offers an accurate and efficient solution for complex computer vision tasks.
• The method's ability to identify adequate accuracy provides computational savings.