Disentangled Geometric Alignment with Adaptive Contrastive Perturbation for Reliable Domain Transfer

2025/05/21 by Collins, Emma, wong, Myungseo, Yun, Kim +2
Computer Science · Engineering · #3D Shape Modeling and Analysis #Computational Geometry and Mesh Generation #Computer Graphics and Visualization Techniques #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences

paper · pdf · doi:10.48550/arxiv.2505.15241

openalex publication_date 2025/05/21 · openalex created_date 2025/10/19 · openalex updated_date 2026/07/30

Abstract

Despite progress in geometry-aware domain adaptation, current methods such as GAMA still suffer from two unresolved issues: (1) insufficient disentanglement of task-relevant and task-irrelevant manifold dimensions, and (2) rigid perturbation schemes that ignore per-class alignment asymmetries. To address this, we propose GAMA++, a novel framework that introduces (i) latent space disentanglement to isolate label-consistent manifold directions from nuisance factors, and (ii) an adaptive contrastive perturbation strategy that tailors both on- and off-manifold exploration to class-specific manifold curvature and alignment discrepancy. We further propose a cross-domain contrastive consistency loss that encourages local semantic clusters to align while preserving intra-domain diversity. Our method achieves state-of-the-art results on DomainNet, Office-Home, and VisDA benchmarks under both standard and few-shot settings, with notable improvements in class-level alignment fidelity and boundary robustness. GAMA++ sets a new standard for semantic geometry alignment in transfer learning.

Citations

• A Theory-Inspired Framework for Few-Shot Cross-Modal Sketch Person Re-Identification
• AIM 2025 Low-light RAW Video Denoising Challenge: Dataset, Methods and Results
• Camel: Energy-Aware LLM Inference on Resource-Constrained Devices
• RobustGS: Unified Boosting of Feedforward 3D Gaussian Splatting under Low-Quality Conditions
• Towards Universal Modal Tracking with Online Dense Temporal Token Learning
• Set You Straight: Auto-Steering Denoising Trajectories to Sidestep Unwanted Concepts
• Decoupled Spatio-Temporal Consistency Learning for Self-Supervised Tracking
• Label Distribution Learning with Biased Annotations by Learning Multi-Label Representation
• EraseAnything: Enabling Concept Erasure in Rectified Flow Transformers
• Phys4DGen: Physics-Compliant 4D Generation with Multi-Material Composition Perception
• Robust Watermarking Using Generative Priors Against Image Editing: From Benchmarking to Advances
• Adversarial Learning for Neural PDE Solvers with Sparse Data
• Cross-Task Attack: A Self-Supervision Generative Framework Based on Attention Shift
• Beyond Dropout: Robust Convolutional Neural Networks Based on Local Feature Masking
• MACE: Mass Concept Erasure in Diffusion Models
• Cross-Modality Perturbation Synergy Attack for Person Re-identification
• ODTrack: Online Dense Temporal Token Learning for Visual Tracking
• Towards Unified Token Learning for Vision-Language Tracking
• Exploiting Multi-Label Correlation in Label Distribution Learning
• TF-ICON: Diffusion-Based Training-Free Cross-Domain Image Composition
• Inaccurate Label Distribution Learning
• Person Re-identification Method Based on Color Attack and Joint Defence
• Eliminate Deviation with Deviation for Data Augmentation and a General Multi-modal Data Learning Method
• Mutual Mean-Teaching: Pseudo Label Refinery for Unsupervised Domain Adaptation on Person Re-identification
• Adversarial Training for Free!
• Bridging Theory and Algorithm for Domain Adaptation
• Theoretically Principled Trade-off between Robustness and Accuracy
• Moment Matching for Multi-Source Domain Adaptation
• Disentangling Adversarial Robustness and Generalization
• Maximum Classifier Discrepancy for Unsupervised Domain Adaptation
• VisDA: The Visual Domain Adaptation Challenge
• Deep Hashing Network for Unsupervised Domain Adaptation
• Towards Deep Learning Models Resistant to Adversarial Attacks
• Virtual Adversarial Training: A Regularization Method for Supervised and Semi-Supervised Learning
• Explaining and Harnessing Adversarial Examples

Related

• GAMA: Geometry-Aware Manifold Alignment via Structured Adversarial Perturbations for Robust Domain Adaptation
• Geometrically Regularized Transfer Learning with On-Manifold and Off-Manifold Perturbation
• Filtered Manifold Alignment
• Unsupervised Domain Adaptation via Discriminative Manifold Embedding and Alignment
• Geometry-aware Domain Adaptation for Unsupervised Alignment of Word\n Embeddings