Vision Foundation Model Adaptation

Tracking Any Point (TAP) has emerged as a fundamental tool for video understanding. Current approaches adapt Vision Foundation Models (VFMs) like DINOv2 via offline finetuning or test-time optimization. However, these VFMs rely on static image pre-training, which is inherently sub-optimal for capturing dense temporal correspondence in videos. To address this, we propose Mask-to-Point (M2P) learning, which leverages rich video object segmentation (VOS) mask annotations to improve VFMs for dense point tracking. Our M2P introduces three new mask-based constraints for weakly-supervised representation learning. First, we propose a local structure consistency loss, which leverages Procrustes analysis to model the cohesive motion of points lying within a local structure, achieving more reliable point-to-point matching learning. Second, we propose a mask label consistency (MLC) loss, which enforces that sampled foreground points strictly match foreground regions across frames. The proposed MLC loss can be regarded as a regularization, which stabilizes training and prevents convergence to trivial solutions. Finally, mask boundary constrain is applied to explicitly supervise boundary points. We show that our weaklysupervised M2P models significantly outperform baseline VFMs with efficient training by using only 3.6K VOS training videos. Notably, M2P achieves 12.8% and 14.6% performance gains over DINOv2-B/14 and DINOv3-B/16 on the TAP-Vid-DAVIS benchmark, respectively. Moreover, the proposed M2P models are used as pre-trained backbones for both test-time optimized and offline fine-tuned TAP tasks, demonstrating its potential to serve as general pre-trained models for point tracking. Code will be made publicly available upon acceptance.

Automated behavior classification is essential for precision livestock farming but faces challenges of high computational costs and limited labeled data. This study systematically compared three approaches: training from scratch (ResNet-18, ViT-Small), frozen feature extraction, and parameter-efficient fine-tuning (PEFT) of the DINOv3 foundation model (6.7 billion parameters). We evaluated QLoRA and DoRA across multiple configurations varying rank (8, 16, 64) and target modules (q_proj versus all-linear layers). With 2,160 verified training images, we assessed generalization of our model on 211,800 test samples, which is essentially a 98:1 test-to-train ratio. Results demonstrated that PEFT substantially outperformed alternatives, where the best QLoRA configuration (all-linear layers and rank=64) achieved 83.16% test accuracy with only 2.72% parameters (183.0M) in 5.8 hours, compared to 72.87% for ResNet-18 (16.8 hours), 61.91% for ViT-Small (18.7 hours), and 76.56% for frozen DINOv3 (17.5 hours). DoRA achieved comparable accuracy (83.14%) but with longer training time (11.0 hours). Notably, increasing adapter capacity consistently improved generalization while simultaneously not causing overfitting: reducing rank from 16 to 8 decreased test accuracy from 78.38% to 77.17%, while expanding from q_proj-only to all-linear layers with rank=64 improved accuracy from 78.38% to 83.16%. This suggests underfitting, instead of overfitting, is the primary challenge when adapting foundation models to agricultural imagery. Our findings provide guidelines for deploying billion-parameter vision models with PEFT in agricultural livestock applications.

Multimodal remote sensing semantic segmentation enhances scene interpretation by exploiting complementary physical cues from heterogeneous data. Although pretrained Vision Foundation Models (VFMs) provide strong general-purpose representations, adapting them to multimodal tasks often incurs substantial computational overhead and is prone to modality imbalance, where the contribution of auxiliary modalities is suppressed during optimization. To address these challenges, we propose MoBaNet, a parameter-efficient and modality-balanced symmetric fusion framework. Built upon a largely frozen VFM backbone, MoBaNet adopts a symmetric dual-stream architecture to preserve generalizable representations while minimizing the number of trainable parameters. Specifically, we design a Cross-modal Prompt-Injected Adapter (CPIA) to enable deep semantic interaction by generating shared prompts and injecting them into bottleneck adapters under the frozen backbone. To obtain compact and discriminative multimodal representations for decoding, we further introduce a Difference-Guided Gated Fusion Module (DGFM), which adaptively fuses paired stage features by explicitly leveraging cross-modal discrepancy to guide feature selection. Furthermore, we propose a Modality-Conditional Random Masking (MCRM) strategy to mitigate modality imbalance by masking one modality only during training and imposing hard-pixel auxiliary supervision on modality-specific branches. Extensive experiments on the ISPRS Vaihingen and Potsdam benchmarks demonstrate that MoBaNet achieves state-of-the-art performance with significantly fewer trainable parameters than full fine-tuning, validating its effectiveness for robust and balanced multimodal fusion. The source code in this work is available at https://github.com/sauryeo/MoBaNet.

Current methods for multimodal medical image classification (M3IC) face two major challenges. First, the prevailing "fusion after unimodal image embedding" paradigm cannot fully exploit the complementary and correlated information in the multimodal data. Second, the scarcity of labeled multimodal medical images, coupled with their substantial domain shift from natural images, impedes the use of cutting-edge Vision Foundation Models (VFMs) for medical image embedding. To jointly address the challenges, we propose a novel Early Intervention (EI) framework. Treating one modality as target and the rest as reference, EI harnesses high-level semantic tokens from the reference as intervention tokens to steer the target modality's embedding process at an early stage. For parameter-efficient VFM adaptation, we introduce Mixture of varied-rank LoRAs (MoR), which employs a small set of LoRA experts with distinct ranks and a bypass-allowed router. Extensive experiments on three public datasets, covering retinal disease recognition, skin lesion recognition, and knee anomaly classification, verify the effectiveness of the proposed method against a number of competitive baselines.

Material classification has emerged as a critical task in computer vision and graphics, supporting the assignment of accurate material properties to a wide range of digital and real-world applications. While traditionally framed as an image classification task, this domain faces significant challenges due to the scarcity of annotated data, limiting the accuracy and generalizability of trained models. Recent advances in vision-language foundation models (VLMs) offer promising avenues to address these issues, yet existing solutions leveraging these models still exhibit unsatisfying results in material recognition tasks. In this work, we propose a novel framework that effectively harnesses foundation models to overcome data limitations and enhance classification accuracy. Our method integrates two key innovations: (a) a robust image generation and auto-labeling pipeline that creates a diverse and high-quality training dataset with material-centric images, and automatically assigns labels by fusing object semantics and material attributes in text prompts; (b) a prior incorporation strategy to distill information from VLMs, combined with a joint fine-tuning method that optimizes a pre-trained vision foundation model alongside VLM-derived priors, preserving broad generalizability while adapting to material-specific features. Extensive experiments demonstrate significant improvements on multiple datasets. We show that our synthetic dataset effectively captures the characteristics of real world materials, and the integration of priors from vision-language models significantly enhances the final performance. The source code and dataset will be released.

Vision transformers (ViTs) - especially feature foundation models like DINOv2 - learn rich representations useful for many downstream tasks. However, architectural choices (such as positional encoding) can lead to these models displaying positional biases and artefacts independent of semantic content. This makes zero-shot adaption difficult in fields like material science, where images are often cross-sections of homogeneous microstructure (i.e. having no preferred direction). In this work, we investigate the positional bias in ViTs via linear probing, finding it present across a range of objectives and positional encodings, and subsequently reduce it by finetuning models to use ALiBi relative positional encoding. We demonstrate that these models retain desirable general semantics and their unbiased features can be used successfully in trainable segmentation of complex microscopy images.

Vision Transformers (ViTs) based vision foundation models (VFMs) have achieved remarkable performance across diverse vision tasks, but suffer from quadratic complexity that limits scalability to long sequences. Existing linear attention approaches for ViTs are typically trained from scratch, requiring substantial computational resources, while linearization-based methods developed for large language model decoders do not transfer well to ViTs. To address these challenges, we propose ViT-AdaLA, a novel framework for effectively adapting and transferring prior knowledge from VFMs to linear attention ViTs. ViT-AdaLA consists of three stages: attention alignment, feature alignment, and supervised fine-tuning. In the attention alignment stage, we align vanilla linear attention with the original softmax-based attention in each block to approximate the behavior of softmax attention. However, residual approximation errors inevitably accumulate across layers. We mitigate this by fine-tuning the linearized ViT to align its final-layer features with a frozen softmax VFM teacher. Finally, the adapted prior knowledge is transferred to downstream tasks through supervised fine-tuning. Extensive experiments on classification and segmentation tasks demonstrate the effectiveness and generality of ViT-AdaLA over various state-of-the-art linear attention counterpart.

Accurate monocular depth estimation is critical in colonoscopy for lesion localization and navigation. Foundation models trained on natural images fail to generalize directly to colonoscopy. We identify the core issue not as a semantic gap, but as a statistical shift in the frequency domain: colonoscopy images lack the strong high-frequency edge and texture gradients that these models rely on for geometric reasoning. To address this, we propose SpecDepth, a parameter-efficient adaptation framework that preserves the robust geometric representations of the pre-trained models while adapting to the colonoscopy domain. Its key innovation is an adaptive spectral rectification module, which uses a learnable wavelet decomposition to explicitly model and amplify the attenuated high-frequency components in feature maps. Different from conventional fine-tuning that risks distorting high-level semantic features, this targeted, low-level adjustment realigns the input signal with the original inductive bias of the foundational model. On the public C3VD and SimCol3D datasets, SpecDepth achieved state-of-the-art performance with an absolute relative error of 0.022 and 0.027, respectively. Our work demonstrates that directly addressing spectral mismatches is a highly effective strategy for adapting vision foundation models to specialized medical imaging tasks. The code will be released publicly after the manuscript is accepted for publication.

Pre-trained image editing models exhibit strong spatial reasoning and object-aware transformation capabilities acquired from billions of image-text pairs, yet they possess no explicit temporal modeling. This paper demonstrates that these spatial priors can be repurposed to unlock temporal synthesis capabilities through minimal adaptation - without introducing any video-specific architecture or motion estimation modules. We show that a large image editing model (Qwen-Image-Edit), originally designed solely for static instruction-based edits, can be adapted for Video Frame Interpolation (VFI) using only 64-256 training samples via Low-Rank Adaptation (LoRA). Our core contribution is revealing that the model's inherent understanding of "how objects transform" in static scenes contains latent temporal reasoning that can be activated through few-shot fine-tuning. While the baseline model completely fails at producing coherent intermediate frames, our parameter-efficient adaptation successfully unlocks its interpolation capability. Rather than competing with task-specific VFI methods trained from scratch on massive datasets, our work establishes that foundation image editing models possess untapped potential for temporal tasks, offering a data-efficient pathway for video synthesis in resource-constrained scenarios. This bridges the gap between image manipulation and video understanding, suggesting that spatial and temporal reasoning may be more intertwined in foundation models than previously recognized

Deploying foundational medical Segment Anything Models (SAMs) via test-time adaptation (TTA) is challenging under large distribution shifts, where test-time supervision is often unreliable. While active test-time adaptation (ATTA) introduces limited expert feedback to improve reliability, existing ATTA methods still suffer from unreliable uncertainty estimation and inefficient utilization of sparse annotations. To address these issues, we propose Evidential Active Test-Time Adaptation (EviATTA), which is, to our knowledge, the first ATTA framework tailored for medical SAMs. Specifically, we adopt the Dirichlet-based Evidential Modeling to decompose overall predictive uncertainty into distribution uncertainty and data uncertainty. Building on this decomposition, we design a Hierarchical Evidential Sampling strategy, where image-wise distribution uncertainty is used to select informative shifted samples, while distance-aware data uncertainty guides sparse pixel annotations to resolve data ambiguities. We further introduce Dual Consistency Regularization, which enforces progressive prompt consistency on sparsely labeled samples to better exploit sparse supervision and applies variational feature consistency on unlabeled samples to stabilize adaptation. Extensive experiments on six medical image segmentation datasets demonstrate that EviATTA consistently improves adaptation reliability with minimal expert feedback under both batch-wise and instance-wise test-time adaptation settings.

Foundation models, pretrained on diverse data, objectives, or modalities for reuse across downstream tasks, are increasingly applied to medical image analysis. Their performance depends not only on pretraining but also on how they are adapted to specific tasks, data regimes, and deployment settings. Whereas previous reviews have considered adaptation within broader syntheses, this review focuses specifically on post-pretraining interventions applied to existing FMs. Using a structured narrative review with a systematic, auditable workflow, we identified 719 eligible reports representing 705 independent studies. Of these, 473 provided controlled comparative evidence: 322 formally published studies formed the primary synthesis and 151 preprint-only studies a sensitivity layer. Adaptation was organized into five non-mutually exclusive families: parameter-level, representation-level, objective-level, data-centric, and architectural/sequence-level adaptation, with model use and update regime analyzed separately. Controlled comparisons showed no consistent ranking among frozen, parameter-efficient, partial, or full updating, or across broader adaptation families; outcomes varied by task, data regime, comparator, and evaluation setting. Clinical-readiness evidence remained limited: 173/705 studies (24.5%) had confirmed independent external validation, 10 (1.4%) reported calibration or uncertainty evidence, and 97/322 formally published comparative studies (30.1%) included a qualifying externally validated arm. Overall, FM adaptation is a context-dependent design space rather than a hierarchy of preferred strategies.

Recently, feature upsampling has gained increasing attention owing to its effectiveness in enhancing vision foundation models (VFMs) for pixel-level understanding tasks. Existing methods typically rely on high-resolution features from the same foundation model to achieve upsampling via self-reconstruction. However, relying solely on intra-model features forces the upsampler to overfit to the source model's inherent location misalignment and high-norm artifacts. To address this fundamental limitation, we propose DiveUp, a novel framework that breaks away from single-model dependency by introducing multi-VFM relational guidance. Instead of naive feature fusion, DiveUp leverages diverse VFMs as a panel of experts, utilizing their structural consensus to regularize the upsampler's learning process, effectively preventing the propagation of inaccurate spatial structures from the source model. To reconcile the unaligned feature spaces across different VFMs, we propose a universal relational feature representation, formulated as a local center-of-mass (COM) field, that extracts intrinsic geometric structures, enabling seamless cross-model interaction. Furthermore, we introduce a spikiness-aware selection strategy that evaluates the spatial reliability of each VFM, effectively filtering out high-norm artifacts to aggregate guidance from only the most reliable expert at each local region. DiveUp is a unified, encoder-agnostic framework; a jointly-trained model can universally upsample features from diverse VFMs without requiring per-model retraining. Extensive experiments demonstrate that DiveUp achieves state-of-the-art performance across various downstream dense prediction tasks, validating the efficacy of multi-expert relational guidance. Our code and models are available at: https://github.com/Xiaoqiong-Liu/DiveUp

Scene-level neural volumetric reconstruction from monocular videos remains challenging, especially under severe domain shifts. Although recent advances in vision foundation models (VFMs) provide transferable generalized priors learned from large-scale data, their scale-ambiguous predictions are incompatible with the scale consistency required by volumetric fusion. To address this gap, we present VFM-Recon, the first attempt to bridge transferable VFM priors with scale-consistent requirements in scene-level neural reconstruction. Specifically, we first introduce a lightweight scale alignment stage that restores multi-view scale coherence. We then integrate pretrained VFM features into the neural volumetric reconstruction pipeline via lightweight task-specific adapters, which are trained for reconstruction while preserving the cross-domain robustness of pretrained representations. We train our model on ScanNet train split and evaluate on both in-distribution ScanNet test split and out-of-distribution TUM RGBD and Tanks and Temples datasets. The results demonstrate that our model achieves state-of-the-art performance across all dataset domains. In particular, on the challenging outdoor Tanks and Temples dataset, our model achieves an F1 score of 73.5 in reconstructed mesh evaluation, substantially outperforming the closest neural volumetric reconstruction competitor, SimpleRecon, with an F1 score of 35.7, as well as the closest VFM competitor, $π^3$, with an F1 score of 67.0.

Accurate visual fault detection in freight trains remains a critical challenge for intelligent transportation system maintenance, due to complex operational environments, structurally repetitive components, and frequent occlusions or contaminations in safety-critical regions. Conventional instance segmentation methods based on convolutional neural networks and Transformers often suffer from poor generalization and limited boundary accuracy under such conditions. To address these challenges, we propose a lightweight self-prompted instance segmentation framework tailored for freight train fault detection. Our method leverages the Segment Anything Model by introducing a self-prompt generation module that automatically produces task-specific prompts, enabling effective knowledge transfer from foundation models to domain-specific inspection tasks. In addition, we adopt a Tiny Vision Transformer backbone to reduce computational cost, making the framework suitable for real-time deployment on edge devices in railway monitoring systems. We construct a domain-specific dataset collected from real-world freight inspection stations and conduct extensive evaluations. Experimental results show that our method achieves 74.6 $AP^{\text{box}}$ and 74.2 $AP^{\text{mask}}$ on the dataset, outperforming existing state-of-the-art methods in both accuracy and robustness while maintaining low computational overhead. This work offers a deployable and efficient vision solution for automated freight train inspection, demonstrating the potential of foundation model adaptation in industrial-scale fault diagnosis scenarios. Project page: https://github.com/MVME-HBUT/SAM_FTI-FDet.git

Although recent Open-Vocabulary Object Detection architectures, such as Grounding DINO, demonstrate strong zero-shot capabilities, their performance degrades significantly under domain shifts. Moreover, many domains of practical interest, such as nighttime or foggy scenes, lack large annotated datasets, preventing direct fine-tuning. In this paper, we introduce Aligned Basis Relocation for Adaptation(ABRA), a method that transfers class-specific detection knowledge from a labeled source domain to a target domain where no training images containing these classes are accessible. ABRA formulates this adaptation as a geometric transport problem in the weight space of a pretrained detector, aligning source and target domain experts to transport class-specific knowledge. Extensive experiments across challenging domain shifts demonstrate that ABRA successfully teleports class-level specialization under multiple adverse conditions. Our code will be made public upon acceptance.

Diabetic retinopathy (DR) is a leading cause of preventable blindness, and automated fundus image grading can play an important role in large-scale screening. In this work, we investigate three CLIP-based approaches for five-class DR severity grading: (1) a zero-shot baseline using prompt engineering, (2) a hybrid FCN-CLIP model augmented with CBAM attention, and (3) a ranking-aware prompting model that encodes the ordinal structure of DR progression. We train and evaluate on a combined dataset of APTOS 2019 and Messidor-2 (n=5,406), addressing class imbalance through resampling and class-specific optimal thresholding. Our experiments show that the ranking-aware model achieves the highest overall accuracy (93.42%, AUROC 0.9845) and strong recall on clinically critical severe cases, while the hybrid FCN-CLIP model (92.49%, AUROC 0.99) excels at detecting proliferative DR. Both substantially outperform the zero-shot baseline (55.17%, AUROC 0.75). We analyze the complementary strengths of each approach and discuss their practical implications for screening contexts.

Modern foundation models provide highly expressive visual representations, yet adapting them to high-resolution medical imaging remains challenging due to limited annotations and weak supervision. Mammography, in particular, is characterized by large images, variable multi-view studies and predominantly breast-level labels, making end-to-end fine-tuning computationally expensive and often impractical. We propose Multiple Instance Learning on Precomputed Features (MIL-PF), a scalable framework that combines frozen foundation encoders with a lightweight MIL head for mammography classification. By precomputing the semantic representations and training only a small task-specific aggregation module (40k parameters), the method enables efficient experimentation and adaptation without retraining large backbones. The architecture explicitly models the global tissue context and the sparse local lesion signals through attention-based aggregation. MIL-PF achieves state-of-the-art classification performance at clinical scale while substantially reducing training complexity. We release the code for full reproducibility.

Zero-shot anomaly detection (ZSAD) requires detecting and localizing anomalies without access to target-class anomaly samples. Mainstream methods rely on vision-language models (VLMs) such as CLIP: they build hand-crafted or learned prompt sets for normal and abnormal semantics, then compute image-text similarities for open-set discrimination. While effective, this paradigm depends on a text encoder and cross-modal alignment, which can lead to training instability and parameter redundancy. This work revisits the necessity of the text branch in ZSAD and presents VisualAD, a purely visual framework built on Vision Transformers. We introduce two learnable tokens within a frozen backbone to directly encode normality and abnormality. Through multi-layer self-attention, these tokens interact with patch tokens, gradually acquiring high-level notions of normality and anomaly while guiding patches to highlight anomaly-related cues. Additionally, we incorporate a Spatial-Aware Cross-Attention (SCA) module and a lightweight Self-Alignment Function (SAF): SCA injects fine-grained spatial information into the tokens, and SAF recalibrates patch features before anomaly scoring. VisualAD achieves state-of-the-art performance on 13 zero-shot anomaly detection benchmarks spanning industrial and medical domains, and adapts seamlessly to pretrained vision backbones such as the CLIP image encoder and DINOv2. Code: https://github.com/7HHHHH/VisualAD

Pathology foundation models learn morphological representations through self-supervised pretraining on large-scale whole-slide images, yet they do not explicitly capture the underlying molecular state of the tissue. Spatial transcriptomics technologies bridge this gap by measuring gene expression in situ, offering a natural cross-modal supervisory signal. We propose MINT (Molecularly Informed Training), a fine-tuning framework that incorporates spatial transcriptomics supervision into pretrained pathology Vision Transformers. MINT appends a learnable ST token to the ViT input to encode transcriptomic information separately from the morphological CLS token, preventing catastrophic forgetting through DINO self-distillation and explicit feature anchoring to the frozen pretrained encoder. Gene expression regression at both spot-level (Visium) and patch-level (Xenium) resolutions provides complementary supervision across spatial scales. Trained on 577 publicly available HEST samples, MINT achieves the best overall performance on both HEST-Bench for gene expression prediction (mean Pearson r = 0.440) and EVA for general pathology tasks (0.803), demonstrating that spatial transcriptomics supervision complements morphology-centric self-supervised pretraining.

Estimating pasture biomass from field images can support grazing decisions without repeated destructive sampling. The task remains difficult because available datasets are small, biomass components are imbalanced and a single photograph contains two spatially distinct views of the sampled quadrat. This work evaluates vision foundation models for five target pasture biomass regression on the 357 image CSIRO dataset. Across 25 model configurations, the analysis covers convolutional and self supervised transformer backbones, seven cross view fusion configurations, models with and without 23 metadata features, classical regressors and frozen foundation features. All comparisons use one locked five fold split and raw scale weighted R^2. DINOv3 Large with cross view gated attention gives the strongest verified single model result, 0.8338 +/- 0.0503 across folds. Under a matched gated depthwise convolution head, replacing DINOv2 Large with DINOv3 Large improves R^2 by 0.0372. Fusion results are not ordered by module complexity. Cross view gated attention is best, while a full Mamba fusion block reaches only 0.7930. Across four matched DINOv3 pathways, metadata narrows the score range from 0.0157 to 0.0013 while changing the mean by only +0.0034. It improves three weaker pathways but lowers the strongest attention model by 0.0061. Classical metadata models are strong, frozen DINOv3 features are linearly useful and Dry Dead biomass remains the hardest target. These results show that backbone choice, cross view interaction and metadata must be tested as separate factors on small agricultural datasets.

Remote sensing image segmentation (RSIS) in federated environments has gained increasing attention because it enables collaborative model training across distributed datasets without sharing raw imagery or annotations. Federated RSIS combined with parameter-efficient fine-tuning (PEFT) can unleash the generalization power of pretrained foundation models for real-world applications, with minimal parameter aggregation and communication overhead. However, the dynamic adaptation of pretrained models to heterogeneous client data inevitably increases update uncertainty and compromises the reliability of collaborative optimization due to the lack of uncertainty estimation for each local model. To bridge this gap, we present FedEU, a federated optimization framework for fine-tuning RSIS models driven by evidential uncertainty. Specifically, personalized evidential uncertainty modeling is introduced to quantify epistemic variations of local models and identify high-risk areas under local data distributions. Furthermore, the client-specific feature embedding (CFE) is exploited to enhance channel-aware feature representation while preserving client-specific properties through personalized attention and an element-aware parameter update approach. These uncertainty estimates are uploaded to the server to enable adaptive global aggregation via a Top-k uncertainty-guided weighting (TUW) strategy, which mitigates the impact of distribution shifts and unreliable updates. Extensive experiments on three large-scale heterogeneous datasets demonstrate the superior performance of FedEU. More importantly, FedEU enables balanced model adaptation across diverse clients by explicitly reducing prediction uncertainty, resulting in more robust and reliable federated outcomes. The source codes will be available at https://github.com/zxk688/FedEU.

Domain Generalization Semantic Segmentation (DGSS) in spectral remote sensing is severely challenged by spectral shifts across diverse acquisition conditions, which cause significant performance degradation for models deployed in unseen domains. While fine-tuning foundation models is a promising direction, existing methods employ global, homogeneous adjustments. This "one-size-fits-all" tuning struggles with the spatial heterogeneity of land cover, causing semantic confusion. We argue that the key to robust DGSS lies not in a single global adaptation, but in performing fine-grained, spatially-adaptive refinement of a foundation model's features. To achieve this, we propose SpectralMoE, a novel fine-tuning framework for DGSS. It operationalizes this principle by utilizing a Mixture-of-Experts (MoE) architecture to perform local precise refinement on the foundation model's features, incorporating depth features estimated from selected RGB bands of the spectral remote sensing imagery to guide the fine-tuning process. Specifically, SpectralMoE employs a dual-gated MoE architecture that independently routes visual and depth features to top-k selected experts for specialized refinement, enabling modality-specific adjustments. A subsequent cross-attention mechanism then judiciously fuses the refined structural cues into the visual stream, mitigating semantic ambiguities caused by spectral variations. Extensive experiments show that SpectralMoE sets a new state-of-the-art on multiple DGSS benchmarks across hyperspectral, multispectral, and RGB remote sensing imagery.

Recent depth foundation models trained on perspective imagery achieve strong performance, yet generalize poorly to 360$^\circ$ images due to the substantial geometric discrepancy between perspective and panoramic domains. Moreover, fully fine-tuning these models typically requires large amounts of panoramic data. To address this issue, we propose RePer-360, a distortion-aware self-modulation framework for monocular panoramic depth estimation that adapts depth foundation models while preserving powerful pretrained perspective priors. Specifically, we design a lightweight geometry-aligned guidance module to derive a modulation signal from two complementary projections (i.e., ERP and CP) and use it to guide the model toward the panoramic domain without overwriting its pretrained perspective knowledge. We further introduce a Self-Conditioned AdaLN-Zero mechanism that produces pixel-wise scaling factors to reduce the feature distribution gap between the perspective and panoramic domains. In addition, a cubemap-domain consistency loss further improves training stability and cross-projection alignment. By shifting the focus from complementary-projection fusion to panoramic domain adaptation under preserved pretrained perspective priors, RePer-360 surpasses standard fine-tuning methods while using only 1% of the training data. Under the same in-domain training setting, it further achieves an approximately 20% improvement in RMSE. The code is available at https://github.com/munimo/RePer360.

Few-shot segmentation (FSS) has gained significant attention for its ability to generalize to novel classes with limited supervision, yet remains challenged by structural misalignment and cross-view inconsistency under large appearance or viewpoint variations. This paper tackles these challenges by introducing VINE (View-Informed NEtwork), a unified framework that jointly models structural consistency and foreground discrimination to refine class-specific prototypes. Specifically, VINE introduces a spatial-view graph on backbone features, where the spatial graph captures local geometric topology and the view graph connects features from different perspectives to propagate view-invariant structural semantics. To further alleviate foreground ambiguity, we derive a discriminative prior from the support-query feature discrepancy to capture category-specific contrast, which reweights SAM features by emphasizing salient regions and recalibrates backbone activations for improved structural focus. The foreground-enhanced SAM features and structurally enriched ResNet features are progressively integrated through masked cross-attention, yielding class-consistent prototypes used as adaptive prompts for the SAM decoder to generate accurate masks. Extensive experiments on multiple FSS benchmarks validate the effectiveness and robustness of VINE, particularly under challenging scenarios with viewpoint shifts and complex structures. The code is available at https://github.com/HongliLiu1/VINE-main.

Source Free Unsupervised Domain Adaptation (SFUDA) is critical for deploying deep learning models across diverse clinical settings. However, existing methods are typically designed for low-gap, specific domain shifts and cannot generalize into a unified, multi-modalities, and multi-target framework, which presents a major barrier to real-world application. To overcome this issue, we introduce Tell2Adapt, a novel SFUDA framework that harnesses the vast, generalizable knowledge of the Vision Foundation Model (VFM). Our approach ensures high-fidelity VFM prompts through Context-Aware Prompts Regularization (CAPR), which robustly translates varied text prompts into canonical instructions. This enables the generation of high-quality pseudo-labels for efficiently adapting the lightweight student model to target domain. To guarantee clinical reliability, the framework incorporates Visual Plausibility Refinement (VPR), which leverages the VFM's anatomical knowledge to re-ground the adapted model's predictions in target image's low-level visual features, effectively removing noise and false positives. We conduct one of the most extensive SFUDA evaluations to date, validating our framework across 10 domain adaptation directions and 22 anatomical targets, including brain, cardiac, polyp, and abdominal targets. Our results demonstrate that Tell2Adapt consistently outperforms existing approaches, achieving SOTA for a unified SFUDA framework in medical image segmentation. Code are avaliable at https://github.com/derekshiii/Tell2Adapt.

In this work, we propose HE-VPR, a visual place recognition (VPR) framework that incorporates height estimation. Our system decouples height inference from place recognition, allowing both modules to share a frozen DINOv2 backbone. Two lightweight bypass adapter branches are integrated into our system. The first estimates the height partition of the query image via retrieval from a compact height database, and the second performs VPR within the corresponding height-specific sub-database. The adaptation design reduces training cost and significantly decreases the search space of the database. We also adopt a center-weighted masking strategy to further enhance the robustness against scale differences. Experiments on two self-collected challenging multi-altitude datasets demonstrate that HE-VPR achieves up to 6.1% Recall@1 improvement over state-of-the-art ViT-based baselines and reduces memory usage by up to 90%. These results indicate that HE-VPR offers a scalable and efficient solution for height-aware aerial VPR, enabling practical deployment in GNSS-denied environments. All the code and datasets for this work have been released on https://github.com/hmf21/HE-VPR.

Geospatial Foundation Models (GFMs) typically lack native support for Hyperspectral Imaging (HSI) due to the complexity and sheer size of high-dimensional spectral data. This study investigates the adaptability of TerraMind, a multimodal GFM, to address HSI downstream tasks without HSI-specific pretraining. Therefore, we implement and compare two channel adaptation strategies: Naive Band Selection and physics-aware Spectral Response Function (SRF) grouping. Overall, our results indicate a general superiority of deep learning models with native support of HSI data. Our experiments also demonstrate the ability of TerraMind to adapt to HSI downstream tasks through band selection with moderate performance decline. Therefore, the findings of this research establish a critical baseline for HSI integration, motivating the need for native spectral tokenization in future multimodal model architectures.

The CLIP model's outstanding generalization has driven recent success in Zero-Shot Anomaly Detection (ZSAD) for detecting anomalies in unseen categories. The core challenge in ZSAD is to specialize the model for anomaly detection tasks while preserving CLIP's powerful generalization capability. Existing approaches attempting to solve this challenge share the fundamental limitation of a patch-agnostic design that processes all patches monolithically without regard for their unique characteristics. To address this limitation, we propose MoECLIP, a Mixture-of-Experts (MoE) architecture for the ZSAD task, which achieves patch-level adaptation by dynamically routing each image patch to a specialized Low-Rank Adaptation (LoRA) expert based on its unique characteristics. Furthermore, to prevent functional redundancy among the LoRA experts, we introduce (1) Frozen Orthogonal Feature Separation (FOFS), which orthogonally separates the input feature space to force experts to focus on distinct information, and (2) a simplex equiangular tight frame (ETF) loss to regulate the expert outputs to form maximally equiangular representations. Comprehensive experimental results across 14 benchmark datasets spanning industrial and medical domains demonstrate that MoECLIP outperforms existing state-of-the-art methods. The code is available at https://github.com/CoCoRessa/MoECLIP.

Training large foundation models from scratch for domain-specific applications is almost impossible due to data limits and long-tailed distributions -- taking remote sensing (RS) as an example. Fine-tuning natural image pre-trained models on RS images is a straightforward solution. To reduce computational costs and improve performance on tail classes, existing methods apply parameter-efficient fine-tuning (PEFT) techniques, such as LoRA and AdaptFormer. However, we observe that fixed hyperparameters -- such as intra-layer positions, layer depth, and scaling factors, can considerably hinder PEFT performance, as fine-tuning on RS images proves highly sensitive to these settings. To address this, we propose MetaPEFT, a method incorporating adaptive scalers that dynamically adjust module influence during fine-tuning. MetaPEFT dynamically adjusts three key factors of PEFT on RS images: module insertion, layer selection, and module-wise learning rates, which collectively control the influence of PEFT modules across the network. We conduct extensive experiments on three transfer-learning scenarios and five datasets in both RS and natural image domains. The results show that MetaPEFT achieves state-of-the-art performance in cross-spectral adaptation, requiring only a small amount of trainable parameters and improving tail-class accuracy significantly.

General vision foundation models (VFMs) have been primarily developed on natural images, and their utility for medical image segmentation is therefore often considered to depend on costly adaptation or domain-specific fine-tuning. In this paper, we revisit this assumption from a different perspective: rather than requiring VFM segmentors to relearn visual regularities, we investigate whether the low-level visual priors necessary for anatomical delineation already lie dormant within general VFMs. We observe that frozen VFMs, despite lacking medical supervision, encode transferable visual regularities. These properties are not exclusive to natural images but are also fundamental to medical image understanding. Motivated by this observation, we propose Latent Universal Medical PriOrs for Segmentation (LUMOS), a novel framework that amplifies general VFM priors to conventional medical segmentors. LUMOS consists of two key components: (1) Pathfinder that distills visual cues from a frozen vision foundation model, and (2) Inspiror that sparks the conventional medical networks with spatial guidance from distilled visual regularities. In this way, the segmentor is relieved from learning complex visual regularities entirely from limited medical annotations and can instead focus on task-specific anatomical delineation. Across diverse medical datasets and token-based VFMs, LUMOS shows that general VFMs can serve as spatial prior generators when their frozen token spaces preserve patch-level pattern relevance. DINO provides stable matched-backbone gains, while SigLIP exposes VFM-specific sensitivity caused by its different token granularity and representation objective.