Learning at a Glance: Towards Interpretable Data-Limited Continual Semantic Segmentation via Semantic-Invariance...

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Continual semantic segmentation (CSS) aims for human-like segmentation models using incremental learning (IL).
• Existing CSS methods struggle with balancing old knowledge retention and new knowledge acquisition.
• Current approaches often require extensive annotated data and lack interpretability.

Purpose of the Study:

• To introduce Learning at a Glance (LAG), an efficient, robust, human-like, and interpretable CSS approach.
• To address the trade-off between knowledge preservation and new learning in incremental settings.
• To reduce the reliance on large-scale annotated data for incremental training.

Main Methods:

• Proposes a model-agnostic architecture, LAG, for efficient CSS.
• Introduces semantic-invariance modeling via semantic feature decoupling (channel-wise and spatial-level).
• Employs asymmetric contrastive learning to preserve semantic-invariant knowledge and enhance robustness.

Main Results:

• LAG achieves competitive CSS efficiency with limited incremental data.
• The semantic feature decoupling effectively balances knowledge inheritance and new term learning.
• Demonstrates superior performance under data-limited conditions using a novel CSS protocol.

Conclusions:

• LAG offers an efficient, robust, and interpretable solution for continual semantic segmentation.
• The proposed semantic-invariance modeling and decoupling methods effectively mitigate catastrophic forgetting.
• LAG advances the state-of-the-art in data-limited continual semantic segmentation settings.