A Survey on Label-Efficient Deep Image Segmentation: Bridging the Gap Between Weak Supervision and Dense Prediction

Area of Science:

• Computer Vision
• Deep Learning
• Image Segmentation

Background:

• Deep learning significantly advances image segmentation, a core computer vision task.
• Current methods require expensive pixel-level annotations, hindering broader application.
• Label-efficient algorithms are crucial for reducing annotation burden in deep image segmentation.

Approach:

• This paper reviews label-efficient image segmentation methods, categorizing them by supervision type (none, inexact, incomplete, inaccurate) and segmentation problem (semantic, instance, panoptic).
• It analyzes how existing methods bridge the weak supervision to dense prediction gap using heuristic priors like cross-pixel similarity, cross-label constraints, cross-view consistency, and cross-image relations.

Key Points:

• A taxonomy organizes methods based on weak label types and segmentation tasks.
• Methods leverage heuristic priors to connect weak supervision with dense predictions.
• The review provides a unified perspective on current label-efficient segmentation techniques.

Conclusions:

• Future research should focus on advancing label-efficient deep image segmentation.
• Developing novel approaches to overcome limitations of weak supervision is essential.
• This review guides future directions in efficient and accurate image segmentation.