Overlap-Aware Feature Learning for Robust Unsupervised Domain Adaptation for 3D Semantic Segmentation

2025-04-02

AI Summary

Unsupervised domain adaptation for 3D point cloud semantic segmentation (PCSS-UDA) suffers from insufficient robustness under realistic environmental perturbations (e.g., rain, fog, snow) and adversarial attacks. This paper identifies two root causes: (i) feature overlap due to misalignment of shared-class region boundaries across domains, and (ii) erosion of target-specific structural patterns caused by over-aggressive domain-invariant learning. Method: We propose an Invertible Attention Alignment Module (IAAM) to suppress feature confusion and a quality-aware contrastive memory bank to enable progressive pseudo-label refinement. Our approach integrates pseudo-label quality-weighted contrastive learning and a novel metric quantifying adversarial and corruption robustness. Contribution/Results: On the SynLiDAR→SemanticPOSS cross-domain benchmark, our method achieves a +14.3% mIoU improvement under adversarial attacks, demonstrating significantly enhanced generalization and stability in complex, real-world perturbation scenarios.

Technology Category

• Computer Vision: Adversarial Attacks & Robustness
• Machine Learning: Adversarial Learning & Robustness
• Multiagent Systems: Adversarial Agents

Application Category

• Search and Retrieval-Augmented AI: Vertical and domain-specific search
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems

Abstract

3D point cloud semantic segmentation (PCSS) is a cornerstone for environmental perception in robotic systems and autonomous driving, enabling precise scene understanding through point-wise classification. While unsupervised domain adaptation (UDA) mitigates label scarcity in PCSS, existing methods critically overlook the inherent vulnerability to real-world perturbations (e.g., snow, fog, rain) and adversarial distortions. This work first identifies two intrinsic limitations that undermine current PCSS-UDA robustness: (a) unsupervised features overlap from unaligned boundaries in shared-class regions and (b) feature structure erosion caused by domain-invariant learning that suppresses target-specific patterns. To address the proposed problems, we propose a tripartite framework consisting of: 1) a robustness evaluation model quantifying resilience against adversarial attack/corruption types through robustness metrics; 2) an invertible attention alignment module (IAAM) enabling bidirectional domain mapping while preserving discriminative structure via attention-guided overlap suppression; and 3) a contrastive memory bank with quality-aware contrastive learning that progressively refines pseudo-labels with feature quality for more discriminative representations. Extensive experiments on SynLiDAR-to-SemanticPOSS adaptation demonstrate a maximum mIoU improvement of 14.3% under adversarial attack.

Problem

• Addresses vulnerability to real-world perturbations in 3D semantic segmentation
• Resolves feature overlap from unaligned boundaries in domain adaptation
• Mitigates feature erosion caused by domain-invariant learning suppressing target patterns

Innovation

• Robustness evaluation model for adversarial resilience
• Invertible attention alignment for domain mapping
• Quality-aware contrastive learning for discriminative representations