Unsupervised domain adaptive segmentation algorithm based on two-level category alignment.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised domain adaptation (UDA) for semantic segmentation often overlooks category information, focusing solely on pixel-level local features.
• This limitation hinders the learning of category-specific inter-domain invariant features, negatively impacting segmentation performance.

Purpose of the Study:

• To develop an effective UDA algorithm that leverages category information for improved semantic segmentation.
• To address the limitations of existing methods by incorporating both image-level and pixel-level category alignment.

Main Methods:

• Proposed UDAca+ algorithm featuring two-level category alignment: image-level using Class Activation Maps (CAM) and pixel-level using pseudo-labels.
• Implemented an adversarial learning strategy in a mixed domain for network training.
• Introduced a confidence calculation method to mitigate issues from noisy image-level pseudo-labels, such as negative transfer and over-alignment.

Main Results:

• UDAca+ effectively captures domain-invariant, category-discriminative feature representations, leading to enhanced segmentation accuracy.
• Achieved state-of-the-art (SOTA) performance on two synthetic-to-real adaptation tasks.
• Demonstrated the algorithm's effectiveness in improving image segmentation within domain adaptation scenarios.

Conclusions:

• The proposed UDAca+ algorithm significantly advances unsupervised domain adaptive semantic segmentation by effectively utilizing category information.
• The two-level category alignment strategy and confidence calculation method provide a robust solution for domain adaptation challenges.