A Likelihood Ratio-Based Approach to Segmenting Unknown Objects.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Out-of-Distribution (OoD) segmentation is crucial for open-world AI perception systems.
• Large foundational models offer robust representations but their OoD capabilities are underexplored.
• Current outlier supervision methods disrupt learned features and are infeasible for large models.

Purpose of the Study:

• To develop an effective outlier supervision method for OoD segmentation in large foundational models.
• To enhance Out-of-Distribution detection without compromising the model's existing feature representations.
• To achieve state-of-the-art performance in detecting unknown objects.

Main Methods:

• Proposed an adaptive, lightweight Unknown Estimation Module (UEM) for outlier supervision.
• UEM learns distributions for outliers and known classes.
• Introduced a likelihood-ratio-based scoring function fusing UEM confidence with inlier network predictions.
• Developed an objective to directly optimize the outlier score.

Main Results:

• Achieved new state-of-the-art performance on multiple Out-of-Distribution segmentation benchmarks.
• Outperformed previous methods by 5.74% in average precision.
• Demonstrated a lower false-positive rate compared to existing approaches.
• Maintained strong inlier segmentation performance.

Conclusions:

• The proposed Unknown Estimation Module (UEM) effectively enhances Out-of-Distribution segmentation.
• This method provides a non-disruptive approach to outlier supervision for large foundational models.
• The approach sets a new standard for robust open-world perception systems.