Learning With Style: Continual Semantic Segmentation Across Tasks and Domains.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Deep learning models struggle with real-world image understanding due to task and domain variability.
• Existing methods address domain adaptation and class incremental learning separately, leaving their unified solution an open problem.
• Continual semantic segmentation under combined task and domain shift presents a significant challenge.

Purpose of the Study:

• To develop a unified framework for continual learning that addresses both domain and task shifts simultaneously.
• To tackle the semantic shift in both input and label spaces within continual learning scenarios.
• To propose a method robust against catastrophic forgetting in evolving environments.

Main Methods:

• Formal introduction of continual learning under task and domain shift.
• Utilizing style transfer techniques to extend knowledge across domains during incremental learning.
• Employing a robust distillation framework to retain task knowledge under incremental domain shift.
• Introducing the Learning with Style (LwS) framework.

Main Results:

• The LwS framework demonstrates the ability to generalize incrementally acquired task knowledge across all encountered domains.
• LwS proves robust against catastrophic forgetting, a common issue in continual learning.
• Extensive experiments on autonomous driving datasets show LwS outperforms existing approaches.

Conclusions:

• The proposed LwS framework offers a robust solution for continual semantic segmentation under combined task and domain shift.
• Existing methods are ill-equipped to handle the complexities of continual learning in dynamic, multi-domain environments.
• LwS advances the field by providing a unified approach to domain and task variability in deep learning.