Unsupervised Learning for Maximum Consensus Robust Fitting: A Reinforcement Learning Approach.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Robust model fitting is crucial for computer vision but challenging with outlier-contaminated data due to computational complexity.
• Existing learning-based methods are often supervised, requiring extensive labeled data.
• There is a need for efficient, unsupervised methods for robust model fitting.

Purpose of the Study:

• To introduce a novel unsupervised learning framework for robust model fitting.
• To develop a method that directly learns to solve robust model fitting without labeled data.
• To create a feature-agnostic approach applicable to various LP-type problems with quasi-convex residuals.

Main Methods:

• Utilizes a Reinforcement Learning (RL) framework, specifically adopting a MaxCon (Maximize Consensus) strategy similar to RANSAC.
• Designs appropriate reward functions, tunable via a parameter, and state encodings for the RL agent.
• Investigates the application of basic and enhanced Q-learning components within the RL framework.

Main Results:

• The proposed unsupervised method demonstrates superior performance compared to existing supervised and unsupervised learning approaches.
• Achieves competitive results against traditional, non-learning-based robust model fitting methods.
• The framework is feature-agnostic and generalizable to a wide range of problems.

Conclusions:

• The novel unsupervised reinforcement learning framework offers an efficient and effective solution for robust model fitting in outlier-contaminated datasets.
• This approach overcomes the limitations of supervised learning by eliminating the need for labeled data.
• The method's generalizability and competitive performance highlight its potential impact on computer vision and related fields.