360SFUDA++: Towards Source-Free UDA for Panoramic Segmentation by Learning Reliable Category Prototypes.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Source-free unsupervised domain adaptation (SFUDA) is crucial for transferring knowledge between domains without source data.
• Pinhole-to-panoramic semantic segmentation faces challenges due to distinct Fields-of-View (FoV), style discrepancies, and panoramic image distortion.

Purpose of the Study:

• To develop an effective SFUDA method for pinhole-to-panoramic semantic segmentation.
• To address semantic mismatches, style discrepancies, and distortion challenges in domain adaptation.

Main Methods:

• Proposed 360SFUDA++ utilizing Tangent Projection (TP) and Fixed FoV Projection (FFP) for knowledge extraction.
• Introduced Reliable Panoramic Prototype Adaptation Module (RPAM) for knowledge transfer at prediction and prototype levels.
• Incorporated Cross-projection Dual Attention Module (CDAM) for feature alignment across projections.

Main Results:

• 360SFUDA++ effectively extracts and transfers knowledge from pinhole models to panoramic domains.
• RPAM and CDAM modules facilitate reliable knowledge adaptation and cross-projection alignment.
• Achieved significantly better performance compared to prior SFUDA methods on synthetic and real-world benchmarks.

Conclusions:

• 360SFUDA++ demonstrates a robust solution for SFUDA in pinhole-to-panoramic semantic segmentation.
• The proposed modules effectively handle domain-specific challenges, leading to state-of-the-art results.
• The method shows strong generalization across diverse indoor and outdoor scenarios.