Chonghua Lv

Chonghua Lv, Dong Zhao, Shuang Wang et al.

Knowledge distillation (KD) has been widely applied in semantic segmentation to compress large models, but conventional approaches primarily preserve in-domain accuracy while neglecting out-of-domain generalization, which is essential under distribution shifts. This limitation becomes more severe with the emergence of vision foundation models (VFMs): although VFMs exhibit strong robustness on unseen data, distilling them with conventional KD often compromises this ability. We propose Generalizable Knowledge Distillation (GKD), a multi-stage framework that explicitly enhances generalization. GKD decouples representation learning from task learning. In the first stage, the student acquires domain-agnostic representations through selective feature distillation, and in the second stage, these representations are frozen for task adaptation, thereby mitigating overfitting to visible domains. To further support transfer, we introduce a query-based soft distillation mechanism, where student features act as queries to teacher representations to selectively retrieve transferable spatial knowledge from VFMs. Extensive experiments on five domain generalization benchmarks demonstrate that GKD consistently outperforms existing KD methods, achieving average gains of +1.9% in foundation-to-foundation (F2F) and +10.6% in foundation-to-local (F2L) distillation. The code will be available at https://github.com/Younger-hua/GKD.

Wei Wang, Dou Quan, Chonghua Lv et al.

Cross-modal remote sensing image (CRSI) registration is critical for multi-modal image applications. However, CRSI mainly faces two challenges: significant nonlinear radiometric variations between cross-modal images and limited textures hindering the discriminative information extraction. Existing methods mainly adopt convolutional neural networks (CNNs) or Transformer architectures to extract discriminative features for registration. However, CNNs with the local receptive field fail to capture global contextual features, and Transformers have high computational complexity and restrict their application to high-resolution CRSI. To solve these issues, this paper proposes RegistrationMamba, a novel Mamba architecture based on state space models (SSMs) integrating multi-expert feature learning for improving the accuracy of CRSI registration. Specifically, RegistrationMamba employs a multi-directional cross-scanning strategy to capture global contextual relationships with linear complexity. To enhance the performance of RegistrationMamba under texture-limited scenarios, we propose a multi-expert feature learning (MEFL) strategy to capture features from various augmented image variants through multiple feature experts. MEFL leverages a learnable soft router to dynamically fuse the features from multiple experts, thereby enriching feature representations and improving registration performance. Notably, MEFL can be seamlessly integrated into various frameworks, substantially boosting registration performance. Additionally, RegistrationMamba integrates a multi-level feature aggregation (MFA) module to extract fine-grained local information and enable effective interaction between global and local features. Extensive experiments on CRSI with varying image resolutions have demonstrated that RegistrationMamba has superior performance and robustness compared to state-of-the-art methods.