Neutral Cross-Entropy Loss Based Unsupervised Domain Adaptation for Semantic Segmentation.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Generalization in semantic segmentation is challenging due to domain shifts.
• Unsupervised domain adaptation (UDA) methods, particularly entropy minimization, are common but have limitations.
• Existing entropy minimization methods can over-sharpen predictions and are biased towards easy samples.

Purpose of the Study:

• To address the limitations of current UDA methods for semantic segmentation.
• To improve generalization performance when source and target domain data distributions differ.

Main Methods:

• Proposed a pixel-level consistency regularization method to introduce a smoothness prior.
• Developed a neutral cross-entropy loss function leveraging consistency regularization.
• Demonstrated an internal neutralization mechanism to mitigate over-sharpness and address gradient bias.

Main Results:

• The proposed method effectively mitigates over-sharpness in prediction distributions.
• The neutral cross-entropy loss inherently tackles the gradient bias towards easy samples.
• Achieved state-of-the-art performance on synthetic-to-real UDA tasks using a lightweight network.

Conclusions:

• The novel approach enhances unsupervised domain adaptation for semantic segmentation.
• Pixel-level consistency regularization and neutral cross-entropy loss are effective for improving generalization.
• The method offers a promising solution for domain shift problems in computer vision.