Semantic segmentation in adverse scenes with fewer labeled images

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Degraded images from adverse scenes pose challenges for high-level vision tasks like semantic segmentation.
• Manual labeling of such images is labor-intensive and impractical for training robust models.

Purpose of the Study:

• To develop a semi-supervised learning approach for semantic segmentation of degraded images requiring significantly fewer labeled samples.
• To address the limitations of traditional semi-supervised methods that demand a high percentage of labeled data.

Main Methods:

• A novel two-step training pipeline is proposed to prevent overfitting by separating labeled and unlabeled image training.
• Introduced a re-parameterization domain adapter (RPDA) for efficient domain adaptation on labeled data.
• Utilized a teacher network for knowledge distillation and a class adaptive threshold with label perception (CATLP) for accurate pseudo-label generation on unlabeled data.

Main Results:

• The proposed method achieves superior performance compared to state-of-the-art methods on public datasets featuring adverse scene images.
• Demonstrated effectiveness with as little as 0.5%-1% labeled images, a substantial reduction from traditional methods.

Conclusions:

• The novel semi-supervised approach significantly alleviates the demand for labeled data in semantic segmentation of degraded images.
• The method offers a practical and efficient solution for domain adaptation in challenging visual environments.