Querying Labeled for Unlabeled: Cross-Image Semantic Consistency Guided Semi-Supervised Semantic Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Semi-supervised semantic segmentation requires accurate pseudo labels for unlabeled data.
• Current methods often overlook the utility of labeled images in pseudo-label generation.
• Leveraging labeled data can enhance the reliability of pseudo labels.

Purpose of the Study:

• To develop a novel approach for semi-supervised semantic segmentation.
• To improve pseudo-label quality by utilizing labeled images for rectification.
• To introduce the Cross-Image Semantic Consistency guided Rectifying (CISC-R) method.

Main Methods:

• Proposes Cross-Image Semantic Consistency guided Rectifying (CISC-R).
• Leverages labeled images to rectify pseudo labels for unlabeled images.
• Estimates pixel-level similarity to create a CISC map for rectification.

Main Results:

• CISC-R significantly improves the quality of generated pseudo labels.
• The method demonstrates superior performance over state-of-the-art techniques.
• Experiments conducted on PASCAL VOC 2012, Cityscapes, and COCO datasets.

Conclusions:

• The CISC-R approach effectively enhances semi-supervised semantic segmentation.
• Explicitly using labeled data for pseudo-label rectification is beneficial.
• The proposed method offers a promising direction for future research.