Adversarial Dense Contrastive Learning for Semi-Supervised Semantic Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Contrastive learning significantly enhances semi-supervised dense prediction tasks like semantic segmentation.
• Existing methods face challenges in selecting effective negative samples and implementing robust data augmentation.

Purpose of the Study:

• To develop an adversarial contrastive learning method tailored for semi-supervised semantic segmentation.
• To address the limitations of negative sample selection and data augmentation in contrastive learning.

Main Methods:

• Adoption of direct learning of adversarial negatives to preserve discriminative past information and boost learning efficiency.
• Implementation of AdverseMix, an advanced data augmentation strategy combining under-performing classes for diverse and challenging samples.
• Utilization of auxiliary labels and classifiers to mitigate the impact of over-adversarial negatives.

Main Results:

• Demonstrated superior performance on Pascal VOC and Cityscapes datasets compared to state-of-the-art methods.
• Achieved significant improvements even with limited labeled data, showcasing the method's efficiency.
• Validated the effectiveness of adversarial negatives and AdverseMix in enhancing segmentation accuracy.

Conclusions:

• The proposed adversarial contrastive learning method offers a significant advancement for semi-supervised semantic segmentation.
• The novel approach effectively overcomes key challenges in negative sampling and data augmentation.
• This work provides a more efficient and robust solution for dense prediction tasks with limited supervision.