Averse Deep Semantic Segmentation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Semantic segmentation assigns labels to pixels, allowing for object identification.
• Errors in segmentation include false positives and false negatives.
• Current methods often use cost matrices for error trade-offs, which may not be intuitive.

Purpose of the Study:

• To develop intuitive methods for controlling specific error types in semantic segmentation.
• To explore absolute constraints on errors rather than relative trade-offs.
• To adapt previous binary classifier approaches for semantic segmentation.

Main Methods:

• Proposing a threshold on the sigmoid layer.
• Modifying gradient descent with a new loss function term.
• Implementing a two-phase training approach.

Main Results:

• The two-phase training method yielded robust results.
• A simple sigmoid layer threshold was sufficient in many scenarios.
• The proposed methods allow for more natural control over segmentation errors.

Conclusions:

• Absolute error constraints offer a more natural approach to semantic segmentation tuning.
• Sigmoid thresholding and two-phase training are effective strategies.
• These methods enhance control over false positives and false negatives in segmentation tasks.