CC4S: Encouraging Certainty and Consistency in Scribble-Supervised Semantic Segmentation.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Deep learning for semantic segmentation requires extensive annotated data.
• Scribble supervision offers a user-friendly annotation alternative but faces challenges with prediction consistency and certainty due to sparse and diverse labels.

Purpose of the Study:

• To propose CC4S, a holistic solution to improve certainty and consistency in scribble-supervised semantic segmentation.
• To address the limitations of sparse and diverse scribble annotations in training deep learning models.

Main Methods:

• CC4S integrates a random walk module for uniform neural representations and a soft entropy loss for deterministic predictions.
• Self-supervision training with a consistency loss on the neural eigenspace encourages consistent predictions.
• A retraining phase with pseudo-labels and a color constraint regularizer further enhances performance.

Main Results:

• CC4S achieves performance comparable to fully supervised semantic segmentation methods.
• The proposed method demonstrates robustness, particularly under extreme supervision conditions.
• Experimental results validate the effectiveness of the random walk module and self-supervision strategy.

Conclusions:

• CC4S offers a robust and effective framework for scribble-supervised semantic segmentation.
• The approach successfully mitigates the challenges posed by sparse and diverse annotations.
• This work advances weakly supervised learning in computer vision by improving model certainty and consistency.