BLPSeg: Balance the Label Preference in Scribble-Supervised Semantic Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Scribble-supervised semantic segmentation offers a low-cost alternative to full annotation.
• Current methods diffuse scribbles using low-level features, but neglect annotation bias.
• Labelers often avoid corners, leading to under-learned regions and incomplete segmentation.

Purpose of the Study:

• To propose BLPSeg, a method to balance annotation bias in scribble-supervised semantic segmentation.
• To improve the completeness and accuracy of segmentation models trained with sparse scribbles.

Main Methods:

• BLPSeg predicts an annotation probability map to identify rare labels.
• A novel BLP loss function up-weights rare annotations to balance training.
• A local aggregation module (LAM) propagates supervision to unlabeled regions.

Main Results:

• BLPSeg effectively balances label preference, leading to more complete segmentation.
• The single-stage BLPSeg method achieves state-of-the-art performance.
• Experimental results demonstrate the significant effectiveness of the proposed approach.

Conclusions:

• BLPSeg successfully addresses the annotation bias challenge in scribble-supervised semantic segmentation.
• The method achieves superior performance compared to existing single-stage and multi-stage approaches.
• BLPSeg offers a promising direction for efficient and accurate semantic segmentation.