Conservative-Progressive Collaborative Learning for Semi-Supervised Semantic Segmentation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Semi-supervised learning for semantic segmentation relies on pseudo-supervision.
• A key challenge is balancing the use of high-quality pseudo-labels versus all available pseudo-labels.

Purpose of the Study:

• To introduce a novel approach, Conservative-Progressive Collaborative Learning (CPCL), to address the pseudo-label tradeoff in semi-supervised semantic segmentation.
• To improve the performance and reliability of semantic segmentation models.

Main Methods:

• CPCL trains two predictive networks in parallel, utilizing both agreement and disagreement between their predictions for pseudo-supervision.
• One network uses intersection supervision with high-quality labels for reliability.
• The other network uses union supervision with all pseudo-labels for exploration, with dynamic loss re-weighting based on prediction confidence.

Main Results:

• CPCL achieves state-of-the-art performance in semi-supervised semantic segmentation tasks.
• The proposed method effectively balances conservative learning from reliable labels and progressive exploration using all labels.

Conclusions:

• CPCL offers an effective strategy for semi-supervised semantic segmentation by intelligently combining different sources of pseudo-supervision.
• The dynamic re-weighting mechanism helps mitigate the impact of potentially noisy pseudo-labels, leading to improved segmentation accuracy.