Revisiting Consistency for Semi-Supervised Semantic Segmentation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning models, particularly for dense prediction tasks like semantic segmentation, heavily rely on large amounts of pixel-level annotated data, which is costly and time-consuming to acquire.
• Semi-supervised learning offers a promising approach to mitigate this data dependency by leveraging unlabeled data.

Purpose of the Study:

• To investigate and optimize semi-supervised learning strategies for efficient deep models in dense prediction tasks.
• To evaluate the impact of different perturbation techniques and consistency enforcement methods on model performance.

Main Methods:

• The study explores semi-supervised algorithms enforcing prediction consistency over perturbed unlabeled inputs.
• A novel competitive perturbation model combining geometric warp and photometric jittering is proposed.
• Experiments focus on efficient model architectures suitable for real-time applications.

Main Results:

• One-way consistency, where only one model instance is perturbed and its gradient is backpropagated, shows clear advantages.
• Perturbing only the 'student' branch within the model architecture further enhances performance.
• The proposed perturbation model, particularly its photometric component, significantly outperforms recent methods.

Conclusions:

• Semi-supervised learning with optimized perturbation strategies is highly effective for dense prediction, reducing reliance on extensive labeled data.
• The proposed approach, especially when applied to efficient models, demonstrates potential to surpass traditional supervised methods using coarse annotations, as evidenced by Cityscapes dataset experiments.