Learning Shape-Invariant Representation for Generalizable Semantic Segmentation

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Supervised semantic segmentation excels but struggles with domain generalization due to domain gaps.
• Existing domain adaptation methods require target data, limiting their use in unavailable domains.
• Domain generalization (DG) aims to train models that perform well on unseen domains without target data.

Purpose of the Study:

• To develop a novel framework for semantic segmentation domain generalization that addresses the domain gap.
• To improve model generalization by learning shape-invariant representations, focusing on object shape discrepancies across domains.
• To enhance semantic segmentation performance in new, unavailable domains.

Main Methods:

• Proposed a Shape-Invariant Learning (SIL) framework to learn shape-invariant representations for better generalization.
• Defined 'structural edge' incorporating object boundary and inner structure for enhanced discrimination.
• Implemented a shape perception learning strategy with texture and structural feature discrepancy losses, and shape deformation augmentation.

Main Results:

• The SIL framework effectively learns shape-invariant representations by implicitly aligning shape distributions at the domain level.
• Experimental results demonstrate state-of-the-art performance for the proposed SIL framework in domain generalization tasks.
• The approach successfully enhances shape perception ability by embedding structural edges as shape priors.

Conclusions:

• The Shape-Invariant Learning (SIL) framework significantly improves domain generalization for semantic segmentation.
• Learning shape-invariant representations is crucial for robust performance across diverse and unseen domains.
• The proposed methods offer a promising direction for addressing domain shift challenges in semantic segmentation.