Vision Foundation Model Adaptation

This paper presents a summary of the Competition on Adapting Foundation Models for Face Recognition Using Synthetic Training Data (AFMFR), held at the 2026 International Joint Conference on Biometrics (IJCB 2026). The competition received a total of eight valid submissions from four distinct teams across two complementary tracks: a Full Data Track, in which participants adapt the CLIP ViT-L/14 foundation model using large-scale synthetic identity data, and a Limited Data Track, designed to reflect more resource-constrained adaptation regimes. All training data was generated exclusively using IDPERTURB. Submitted solutions are ranked based on verification and identification performance across a diverse suite of benchmarks, including LFW, CFP-FP, AgeDB-30, CALFW, CPLFW, IJB-B, IJB-C, and TinyFace, using the Borda count method. Fairness evaluation is additionally conducted on the RFW dataset across four demographic groups. The results demonstrate that adaptation of the CLIP foundation model with synthetic training data substantially improves over the off-the-shelf model and, in several cases, surpasses the baseline. Notably, full fine-tuning with Sub-Center ArcFace (DMSTI-Neurotechnology) leads the Full Data Track, while rank-stabilized LoRA adaptation (Idiap-BSP) proves most effective under limited-data conditions.

Efficient surgical segmentation empowers clinical diagnosis, intraoperative monitoring, and downstream robotic pipelines for reconstruction and simulation. Although prompt-driven foundation models like Segment Anything Model 3 (SAM3) achieve strong segmentation performance on natural images, surgical data exhibits domain gaps against its pre-training data, resulting in degraded segmentation accuracy. Furthermore, existing medical SAM methods require full-parameter fine-tuning, incurring heavy computational consumption and low efficiency. To address these limitations, this work proposes a parameter-efficient Low-Rank Adaptation (LoRA) adaptation of SAM3 for surgical concept segmentation. We inject low-rank adapters into the prompt encoder, detector and tracker while fully freezing the vision backbone, which only optimizes 0.98% of the total model parameters and supports training on a single consumer GPU. Comprehensive experiments demonstrate that our method consistently outperforms zero-shot SAM3 and other mainstream baselines, and the generated segmentation results can be directly deployed to support downstream robotic surgical scene reconstruction and physical simulation pipelines.

Vision foundation models are increasingly reused as frozen backbones for downstream visual recognition, making parameter-efficient adaptation a central problem. Prompt-based adaptation, including Visual Prompt Tuning (VPT), provides a lightweight way to specialize these models, but its layer-wise behavior remains poorly understood: performance is sensitive to prompt depth, placement, and task distribution, and gains on standard in-domain benchmarks do not always translate into robust generalization. We argue that this limitation is not solely an optimization issue, but a layer-wise information allocation issue: existing prompt-based methods lack principled control over what prompt-conditioned representations should preserve, suppress, and propagate across depth. Inspired by the Information Bottleneck principle, we introduce Prompted Information Bottlenecks (PIB), a framework that regularizes layer-wise compression-sufficiency trade-offs and promotes a more coherent cross-layer information path. The key idea is that effective adaptation should be minimal yet sufficient, retaining task-relevant local evidence in earlier layers while progressively discarding nuisance factors and redundant details in deeper layers. Extensive experiments show that PIB achieves strong performance across 34 datasets, reaching 92.1% on FGVC, 93.01% on HTA, and 77.33% on VTAB-1k, while tuning only 0.35% parameters on average across the main settings. Beyond benchmark accuracy, PIB helps explain the non-monotonic behavior of prompt capacity scaling, reduces shortcut reliance, and improves robustness under distribution shift and fine-grained recognition settings. These results position PIB as both a practical method and an information-allocation perspective for adapting frozen vision foundation models. Our code is available at https://github.com/itsnotacie/MM-26-PIB

With the increasing diversity of spoofing attacks, there is a growing demand for unified Face Anti-Spoofing (FAS) models capable of detecting both physical and digital threats. While existing Vision-Language Models (VLMs) demonstrate high generalization in this context, they heavily rely on complex multimodal fusion and external text encoders. In this paper, we propose DINO-VPT, a lightweight, vision-only framework leveraging hierarchical visual prompt tuning. By dynamically injecting prompts conditioned on input features via a Prompt Routing Network (PRN), our method effectively disentangles diverse spoofing artifacts without requiring multimodal fusion. Evaluations on the UniAttackData benchmark demonstrate that DINO-VPT achieves higher accuracy than state-of-the-art VLM-based methods. Our results indicate that a properly structured vision-only architecture can achieve state-of-the-art performance in unified FAS without the need for multimodal supervision.

Vision-based automation is an excellent candidate for reducing manual labor in greenhouse crop production and phenotyping. However, progress is constrained by the lack of annotated training data. Recent advances in vision-based foundational models have shown promising results in zero-shot generalization to novel domains, but their performance drops in complex agricultural environments. In this work, we present a sim-to-real framework for tomato plant segmentation that combines synthetic data generation with fine-tuning of a foundation model. We model a commercial cherry tomato greenhouse and use it to generate a large-scale synthetic dataset under diverse viewpoints, lighting conditions, and plant morphology. Subsequently, we fine-tune the Segment Anything Model 3 (SAM 3) on the synthetic dataset, specializing its text-conditioned segmentation behavior for greenhouse crop organs while retaining the general visual prior that makes zero-shot transfer possible. By evaluating our framework on multiple real-world greenhouse datasets, we demonstrate that combining synthetic data with SAM 3 fine-tuning significantly improves segmentation performance and model confidence. To support community benchmarking, we publicly release the procedural model, the generated synthetic dataset, and our fine-tuned SAM 3 weights.

Recent geometric foundation models (e.g., Metric3D, Depth Anything and UniDepth) have substantially improved monocular depth estimation (MDE) in both cross-scene generalization and metric-scale prediction, yet these gains have not translated to tiny models. We bridge this gap with DepthART (Depth Anything Rethought for Tiny Models), which is a compact MDE model for on-device deployment across diverse scenes. We first identify two capacity-driven bottlenecks in tiny models: (i) overfitting to dataset-specific distribution bias and (ii) unstable metric adaptation under camera shift, where full fine-tuning easily damages transferable geometry. Accordingly, DepthART combines two simple but effective strategies: a bias-resistant data sampling scheme to reduce distribution bias under the same training budget, and a camera-conditioned fine-tuning protocol that freezes the distilled encoder and adjusts metric scale conditioned on intrinsics while better preserving cross-dataset generalization. Across datasets, DepthART consistently surpasses previous tiny baselines in both zero-shot generalization and metric accuracy (e.g., zero-shot $δ_1$=0.964 for DepthART-S on NYUD v2), and in some cases approaches heavy models. We further provide a scalable model family, with DepthART-S reaching 347/245 FPS (strict FP32) on an RTX A6000 at $224^2/448^2$, 102 FPS (TF32) on a Orin NX 8GB, and over 15 FPS (FP32) on a Jetson Nano 4GB.

Mild Cognitive Impairment is a critical early stage of cognitive decline that frequently precedes Alzheimer's disease, yet its automated detection from neuropsychological drawing tests remains fundamentally constrained by data scarcity, class imbalance, and diagnostic ambiguity near clinical boundaries. Existing methodologies attempt to bypass these constraints using computationally expensive, fully fine-tuned hybrid architectures that relegate spatial explainability to a post-hoc approximation rather than an intrinsic model property. We propose a parameter-efficient framework utilizing frozen DINOv2-Small model adapted via three modality-specific learnable prompt tokens while Operating with 1.19 million trainable parameters, each token serves as a query in a shared cross-attention layer over the source image patch tokens. Crucially, spatial explainability is achieved directly through these attention maps; as a structural consequence of the architecture. Then task-conditioned embeddings fused via an attention module to quantify modality-level importance per subject. To handle boundary ambiguity, a MoCA-adapted focal loss introduced that integrates continuous cognitive scores into the training target, loss modulation, and adaptive sample weighting, strictly generalizing standard soft-label approaches. Under stratified five-fold cross-validation, the proposed architecture yields an MCI-class F1 of 0.641 and an AUC of 0.795, outperforming the computationally heavier ResViT baseline by 0.110 in MCI-class F1.

Defect segmentation in additive manufacturing (AM) X-ray computed tomography (XCT) images remains challenging due to severe class imbalance and large distribution shifts across scan conditions. Although recent foundation models such as the Segment Anything Model (SAM) provide strong general-purpose segmentation priors, their natural-image pre-training transfers poorly to the AM XCT domain, where defects appear as subtle non-semantic microstructural anomalies. Moreover, adapting SAM to the AM domain is further limited by the large domain gap and scarcity of labeled real XCT data. We present XCT-SAM, a sequential parameter-efficient adaptation framework for AM XCT defect segmentation. Instead of adapting SAM directly from natural images to XCT data, we first fine-tune Conv-LoRA adapters on an alloy-microstructure dataset and subsequently transfer the adapted model to XCT images, progressively bridging the domain gap. Using Conv-LoRA adapters with rank r=2, the framework injects convolutional spatial inductive bias into SAM's backbone while training approximately 4.15M parameters and keeping over 99% of the model frozen. We evaluate XCT-SAM on out-of-distribution CycleGAN-XCT benchmarks and real-world NIST XCT scans. Across both settings, XCT-SAM consistently outperforms zero-shot SAM and other domain-adapted SAM baselines, achieving the best overall IoU and Dice scores. These results demonstrate the effectiveness of intermediate domain adaptation with parameter-efficient adapters for industrial XCT defect segmentation. The source code is publicly available at https://github.com/Mahedi-61/XCT-SAM.git

Foundation models (FMs) have transformed machine learning from isolated task-specific model development toward general-purpose models pretrained on broad data and adapted to multiple downstream tasks. Earth observation (EO) is an important domain for this paradigm because satellite and airborne archives are large, high-revisit, and increasingly multimodal, while reliable field labels are often sparse. Remote sensing foundation models (RSFMs) cannot be transferred reliably/optimally without domain-specific adaptation. This is because EO data are governed by measurement physics and operational decision constraints. This chapter reviews the design principles arising from these domain-specific constraints. It first defines the FMs paradigm in remote sensing (RS), then synthesizes the current model landscape, pretraining objectives, architecture designs, downstream adaptation and trustworthiness requirements. The chapter also incorporates recent benchmark evidence showing that no single geospatial foundation model is universally best and that inconsistent evaluation remains a major issue to fair comparison and reliable deployment. In addition, two brief environmental monitoring case studies; physics-informed spectral targeted masking for harmful algal bloom prediction and reinforcement learning for adaptive environmental monitoring station selection to illustrate the FMs domain-guided principles in practice. This chapter posits that next-generation RSFMs should be evaluated not only by benchmark accuracy, but also by modality-aware transfer and physically plausible representations for trustworthy EO decisions.

Pre-trained Vision Foundation Models (VFMs) provide strong visual representations for diverse downstream tasks. The key challenge of VFM adaptation stems from the prohibitive costs of full fine-tuning and catastrophic forgetting. To address this, Low-Rank Adaptation (LoRA) has emerged as the prevailing paradigm for Parameter-Efficient Fine-Tuning (PEFT). However, LoRA is typically designed for transformer self-attention layers parameterized by 2D matrices. Since convolutional kernels inherently couple spatial and channel information within a 4D tensor, forcing them into a monolithic 2D matrix disrupts the inherent spatial topology. In this paper, we propose Low-Rank Convolutional Adaptation (LoCA), a convolution-aware PEFT framework that addresses spatial-channel entanglement by decoupling channel and spatial adaptation. LoCA introduces a low-rank channel adaptation for dense cross-channel mixing and refines spatial bases extracted from pre-trained kernels via Singular Value Decomposition (SVD). Experimental results show that LoCA preserves pre-trained spatial priors and achieves competitive or state-of-the-art performance across fine-grained classification, domain-generalized semantic segmentation, and generative benchmarks.

Accurate breast cancer classification from mammography requires effective integration of complementary information from craniocaudal (CC) and mediolateral oblique (MLO) views, which provide a more complete characterization of breast abnormalities. However, existing multi-view learning approaches typically rely on feature-level aggregation or single-stage cross-attention, which can entangle view-specific and shared representations and restrict interaction to limited network depths. To address these limitations, we propose a token-centric dual-view learning framework that unifies prompt-based adaptation and cross-view fusion within a frozen vision transformer backbone. The framework reformulates inter-view interaction as structured token-level communication, where dedicated fusion tokens explicitly encode bidirectional information exchange between CC and MLO views via cross-attention, serving as intermediate carriers of cross-view dependencies rather than relying on direct feature fusion. Unlike conventional methods that apply fusion at a single layer, fusion modules are inserted at multiple transformer depths, enabling progressive and repeated interaction across the encoder hierarchy. Fusion tokens are reintegrated into the token sequence and refined by subsequent transformer layers, facilitating hierarchical propagation of complementary information while preserving view-specific structure. Experiments on VinDr-Mammo and CMMD datasets demonstrate consistent improvements over linear probing, prompt-only adaptation, and conventional fusion baselines. On the VinDr-Mammo BI-RADS classification task, the framework achieves 50.40% F1-score and 0.8090 AUC, including a 0.10 AUC improvement over a dual-view fusion baseline in the binary setting. Ablation studies further validate the effectiveness of token-based fusion and multi-depth interaction design.

Adversarial robustness in Unsupervised Domain Adaptation (UDA) remains a significant challenge due to noisy pseudo labels and inherent distributional shifts between the clean source and adversarially perturbed target domains. Existing approaches often fail to achieve an optimal trade-off between robustness and accuracy, as pseudo-labels generated by domain-adapted models tend to introduce classification errors under adversarial attacks. In this work, we propose \textbf{SFT+RL}, a two-stage robust UDA framework that integrates Supervised Fine Tuning (SFT) and Reinforcement Learning (RL) on top of CLIP's pre-trained visual encoder. In the SFT stage, we adversarially fine-tune a linear classifier using PGD-based perturbations over the labelled source domain while partially unfreezing CLIP's projection layer. It allows adaptation to adversarial noise while preserving CLIP's rich semantic priors. We introduce a confidence-guided pseudo-labeling strategy in the RL stage to annotate unlabeled target samples progressively. Pseudo labels are filtered using a decaying confidence threshold to balance quality and coverage, and the model is trained on a composite dataset formed by combining clean source samples with high-confidence target samples. Adversarial training is applied to mixed batches of clean and adversarial examples to enhance cross-domain robustness. Comprehensive evaluations on three benchmark datasets OfficeHome~\cite{tomm-ude}, PACS~\cite{pacs}, and VisDA~\cite{visda} demonstrate the effectiveness of our approach. Notably, \textbf{SFT+RL} achieves average improvements of \textbf{10.2%} in clean accuracy and \textbf{15.8%} in adversarial robustness across all three datasets, outperforming existing state-of-the-art methods.

Recent 3D foundation models, such as DUSt3R, MASt3R, VGGT, $π^3$, and Depth Anything 3, provide strong feed-forward depth and pose estimates on pinhole imagery, but degrade sharply under fisheye camera geometry. We show that this failure is partly caused by a pinhole camera bias in the positional encodings of pretrained 3D foundation models, and propose RayTun3R, a lightweight camera adaptation approach. It keeps the pretrained network fixed and adapts only lightweight components tied to token position and camera geometry. RayTun3R learns parameter-efficient residual corrections to absolute and rotary positional encodings, together with parameter-free tokenization and corrections to prediction-grid coordinates that remove residual pinhole assumptions. The resulting adapter contains only 10,752 trainable parameters and can be learned from a short temporal segment using geometric losses. Once adapted, RayTun3R transfers effectively to the remaining frames of the sequence without incurring additional runtime costs. Across diverse fisheye datasets with fields of view from $110^\circ$ to $200^\circ$, our adapter reduces rotation error by $2$-$12\times$ relative to the unadapted model, outperforms LoRA while using $\sim\!14\times$ fewer trainable parameters, improves pose over adaptation-free baselines while avoiding their multi-view inference cost, and remains competitive on depth accuracy.

Adapting frozen vision foundation models to fine-grained segmentation now largely depends on backbone selection. Whether the backbone applies global attention to a high-resolution token set predicts whether a low-rank adapter turns resolution into accuracy. Isotropic ViTs attend globally over the full grid and keep improving with resolution; hierarchical backbones confine early attention to local windows and pool the grid before their global stages, plateauing at lower resolutions. A controlled six-backbone study establishes the pattern, and editing the backbone points to the cause: pooling keeps the benefit, removing global attention does not. The effect is specific to low-rank adaptation. Under one fixed pipeline, SALT (Side-stem, Attention-gated U-Net, Low-rank Tuning), one RGB-only pass on a strong isotropic backbone wins the best S-measure on the four data-matched camouflaged sets, and leads every marine and salient set. It reaches a new state of the art on both marine-animal benchmarks (MAS3K mIoU 0.878).

Modern pretrained vision models achieve strong accuracy but demand substantial GPU memory for fine-tuning, making edge deployment impractical. This paper compares five parameter-efficient fine-tuning (PEFT) methods (Full FT, LoRA, AdaLoRA, QLoRA, BitFit) on Transformers- (ViT-Small, TinyViT) and Mamba-based vision backbones (Vim-Small, MambaVision-T) under an on-device VRAM budget (e.g., 2 GB), together with three gradient-checkpointing strategies (none, static, and a proposed memory-budget-aware adaptive algorithm); and we evaluate three families of foundation-model baselines: zero-shot contrastive vision language models (OpenCLIP, SigLIP), self-supervised vision backbones with lightweight evaluation protocols (DINOv2), and autoregressive VLMs for prompt-based classification (PaliGemma, MobileVLM, SmolVLM). Experiments on CIFAR-100 and DTD report accuracy, training time, energy, and the NetScore family of multi-objective metrics, which we extend with two deployment-aware variants. QLoRA and BitFit cut energy 20-30% at a 1-2% accuracy cost; the adaptive algorithm reduces peak memory 43-79% with 9-30% energy overhead. DINOv2 surpasses fine-tuned models on CIFAR-100 (0.917 vs. 0.897) at a fraction of the energy, while small autoregressive VLMs remain uncompetitive.

Zero-shot object counting (ZOC) aims to count instances of arbitrary object categories specified only through textual prompts. Recent training-free approaches leverage foundation models such as SAM to reformulate counting as a prompt-driven segmentation task, eliminating the need for costly counting-specific training data with point-level annotations. More recently, SAM3 introduced promptable concept segmentation, enabling the zero-shot segmentation of all instances corresponding to a text-defined concept. However, SAM3 struggles in densely populated scenes containing numerous small objects, where limited image resolution and insufficient attention to target-relevant regions often lead to missed instances and poor instance separation, hindering accurate object counting. To address this limitation, we propose AdaCount, a training-free framework for ZOC based on similarity-guided spatial and feature adaptation. AdaCount first estimates a prototype-driven similarity map that identifies target-relevant regions. This similarity map subsequently guides two complementary adaptations: (i) similarity-guided spatial warping, which reallocates image resolution toward target instances, and (ii) feature modulation, which amplifies target-relevant encoder representations. Together, these adaptations enable SAM3 to devote greater representational capacity to target-relevant regions while preserving global image context, without requiring any model retraining. Extensive experiments across six diverse counting benchmarks establish AdaCount as a new SOTA among training-free ZOC approaches.

In modern high-throughput industrial production lines, product configurations and visual characteristics frequently change, making it impractical to collect and annotate data for every new scenario. This dynamic setting makes Zero-Shot Anomaly Detection (ZSAD) particularly suitable, as it enables defect detection without requiring training on target-specific samples. Although recent ZSAD approaches show promising results, they are computationally intensive and thus unsuitable for deployment on resource-constrained devices. We propose LiZAD: a lightweight framework designed for real-time ZSAD specifically tailored for use on edge devices. The proposed approach pairs the dense and spatially aware visual features of DINOv3, crucial for precise pixel-level localization, with the highly computationally efficient text embeddings of MobileCLIP2. These features are then mapped into a shared latent space via low-memory trainable projection heads. Compared to six state-of-the-art ZSAD models, LiZAD achieves an average memory reduction of 61.5%, a parameter reduction of 74.6%, and a speedup of 3.02x in terms of latency. Despite substantial reductions in computational and memory costs, our approach maintains competitive anomaly detection performance, dropping the average P-AUROC by just 6.4% relative to the best state-of-the-art model across the VisA, BTAD, MPDD, and MVTec-AD datasets. Finally, it is successfully deployed on the NVIDIA Jetson NX and Jetson AGX edge devices and tested on the real production line of the Industrial Computer Engineering Laboratory (ICE Lab) at the University of Verona. The code is available at https://github.com/intelligolabs/LiZAD.

Despite the intense engagement surrounding low-level vision generalist models, their effectiveness in zero/few-shot scenarios beyond learned tasks remains unverified. The primary challenge of developing an ideal generalist lies in achieving the ability to generalize from new unseen tasks, which also can be assessed by matched quantitative criteria. Existing methods have made some progress in prompt engineering but have not systematically explored this gap across a wide range of low-level visual tasks. Stimulated by the problem, we propose Hidden-Shot, an implicit prompt mechanism aimed at exploring low-level task adaptation in a vision generalist model. Specifically, the method extracts implicit visual task-based information, utilizes a global task-aware textural prompt, and selectively merges implicit information with in-task processing information to enhance one-shot capabilities in new tasks. The overall design performs direct injection in a cost-effective manner, while minimally altering the architecture of the original generalist model. Additionally, we introduce a data-driven evaluation framework termed C/U assessment to cover two basic scenarios, 3C4U (3 conventional and 4 unconventional tasks) for retraining existing models and 3C7U (3 conventional and 7 unconventional tasks) for training from scratch, as a comprehensive assessment to systematically test the generalization ability of low-level generalist models. Experiments on seven and ten datasets outperform the state-of-the-art vision generalist model, respectively verified by 3C4U and 3C7U framework. Our presented Hidden-Shot approach demonstrates superior performance on one-shot new tasks while maintaining consistent performance on existing tasks.

Foundation video diffusion models are increasingly viewed as world simulators for embodied agents, yet their pretraining on internet-scale generic video leaves them poorly aligned with real-world deployment domains. We study parameter-efficient adaptation of a pretrained foundation video world model to retail scenes: when synchronized egocentric and exocentric video of the same activity are available, which viewpoint of training data produces the strongest adapted model? We introduce RetailSMV (Retail Synchronized Multi-View), a corpus of 32,105 captioned retail clips from five supermarkets with synchronized ego/exo capture from the store-staff perspective (stocking, arranging, weighing, managing supply carts, scanning at checkout), rather than the customer-centric framing of prior retail video corpora, and train three matched Low-Rank Adaptation (LoRA) configurations of Cosmos3-Nano (egocentric-only, exocentric-only, combined) under identical hyperparameters. On a 200-clip held-out test set evaluated with seven complementary metrics under a strict paired statistical protocol, exocentric-only adaptation matches or exceeds combined adaptation on six of seven point estimates and is significantly better on LPIPS, PSNR, and DreamSim, despite training on only 15,985 exocentric clips (versus 32,105 for combined). A symmetric paired comparison further shows that adding exocentric data to egocentric-only training helps while adding egocentric data to exocentric-only training hurts. The absolute adaptation gap is largest at the shortest rollout time, identifying the near-horizon prediction window as the regime in which adaptation is most beneficial.

Low-resolution face recognition (LR-FR) remains a challenging task due to poor feature extraction and aggregation, as probe images often contain limited identity information resulting from extreme degradations such as blur, occlusion, and low contrast. Additionally, the domain gap between high-resolution (HR) gallery images and low-resolution (LR) probe images poses a significant challenge. A single feature encoder struggles to generalize effectively across both domains when fine-tuned on an LR dataset, and this issue is further magnified by catastrophic forgetting. To address these challenges, we propose FaceMoE, an effective adaptation of Mixture of Experts (MoE) transfomer architecture for low-resolution face-recognition . Specifically, we introduce multiple specialized feed-forward network (FFN) experts and incorporate a top-k router, which dynamically assigns tokens to appropriate experts. This design emergently promotes specialization across experts for different semantic regions of the face, which enables FaceMoE to perform resolution-aware feature extraction. Moreover, the top-k router facilitates sparse expert activation, enabling the model to preserve pretrained knowledge when finetuned on a LR dataset, while increasing model capacity without proportional computational overhead. FaceMoE is trained with a combined face recognition loss, router z-loss, and load balancing loss to ensure expert specialization and stable training. To the best of our knowledge, this is the first work leveraging MoE for LR-FR. Extensive experiments across eleven datasets, spanning HR, mixed-quality, and LR benchmarks, demonstrate that FaceMoE significantly outperforms state-of-the-art methods. Code: https://github.com/Kartik-3004/FaceMoE

This paper presents a safety-centered empirical evaluation of uncertainty-aware last-layer adaptation for referable diabetic retinopathy screening using RETFound, a self-supervised vision-transformer retinal foundation model used here as a frozen feature encoder, and the public APTOS 2019 and DDR diabetic retinopathy fundus image datasets. We compare a cached-feature softmax head, post-hoc temperature scaling, variational Bayesian last-layer heads, a diagonal Laplace last-layer approximation, and an SNGP-style cached-feature head. On APTOS, uncertainty-aware operating points improved sensitivity and selective-referral behavior. The strongest APTOS selective-referral result deferred approximately 20 percent of cases and reduced accepted-case false negatives to zero while preserving high accepted-case specificity. However, threshold tuning also reduced false negatives at high false-positive cost, so false-negative reduction alone was not unique to Bayesian modeling. On DDR, native Bayesian heads qualitatively reproduced the APTOS direction but with weaker tradeoffs, while the APTOS-trained SNGP checkpoint transferred poorly and failed to provide useful external selective-referral behavior. These results highlight the value of safety-centered evaluation beyond aggregate accuracy: uncertainty-aware last-layer heads can improve internal safety-oriented operating points, but trustworthy retinal screening claims require explicit safety-coverage evaluation and second-dataset validation under shift.

Referring Remote Sensing Image Segmentation (RRSIS) seeks to localize and segment the target object or region specified by a natural language expression in a remote sensing image. While existing RRSIS models have benefited from large-scale foundation models, they predominantly rely on full fine-tuning. These approaches are computationally intensive and may weaken the generalization ability of pre-trained models, as extensive fine-tuning on significantly smaller downstream datasets can distort the well-structured feature representations learned during large-scale pre-training. Although Parameter-Efficient Tuning (PET) offers a potential alternative, existing PET frameworks primarily focus on single-modal optimization, failing to capture the complex cross-modal dependencies required for multimodal reasoning, while simultaneously struggling to bridge the substantial domain gap between natural scenes and aerial imagery. To address these limitations, we propose a novel framework, Semantic-driven Scale and Spatial Selection for Efficient Cross-modal Alignment (S4ECA), which enables effective and efficient cross-modal interaction through parameter-efficient adaptation. Specifically, we design a dual-encoder adapter architecture. The textual adapter employs learnable queries to distill highly semantic language proxies from word-level embeddings, facilitating early grounding. Simultaneously, the visual adapter refines hierarchical feature representations through a multi-scale dense extractor, followed by a language-guided scale and spatial selection mechanism that dynamically emphasizes relevant visual contexts, ensuring precise cross-modal alignment. By updating only 2.4% of the backbone parameters, our proposed model achieves state-of-the-art performance on the RRSIS-D and RefSegRS datasets, demonstrating superior efficiency and precision in complex aerial scenarios.

Automated image recognition is increasingly used to scale ecological monitoring beyond manual annotation, yet ecologists lack evidence-based guidance on how much labelling effort reliable deployment at new sites requires. We present a decision framework quantifying the trade-off between labelling effort and recognition accuracy when transferring vision systems across marine habitats. The benchmark spans five datasets, three oceans, and three taxonomic groups (fish, corals, invertebrates), from tropical reefs in the Great Barrier Reef and French Polynesia to a temperate Danish fjord. We evaluated four recognition models (DINOv2, CLIP, ResNet-50, EfficientNet-B4) under four adaptation strategies (linear probing, LoRA, Visual Prompt Tuning, full fine-tuning) across three protocols: within-habitat transfer across 20 reef sites (240 runs), cross-dataset geographic transfer along a difficulty gradient (40 runs), and few-shot adaptation curves with 0-100 labelled samples per class (648 runs). Frozen self-supervised foundation features (DINOv2 + linear classifier, 1,538 trainable parameters) generalised to unseen reef sites at least as well as fully fine-tuned convolutional baselines four orders of magnitude larger; they learned species-diagnostic, habitat-invariant representations, whereas baselines encoded habitat-specific shortcuts that fail at new sites. As few as 10-20 labelled images per species sufficed to deploy reliable recognition at a new site, cutting annotation effort by roughly an order of magnitude. Solution. Programmes expanding to new sites can deploy reliable recognition by pairing a frozen, open foundation model (DINOv2) with a simple linear classifier and annotating only 10-20 images per species - roughly 1-4 hours per site. The framework lets programmes budget labelling effort against expected accuracy across sites, ecosystems, and platforms.

In real-world deployments, scene text detectors inevitably face distribution shifts beyond the training distribution. Prior work often depends on large-scale scene-text pretraining, yet evaluation under cross-domain changes and real-world imaging degradations remains limited. We propose TextDS, an efficient framework for scene text detection under distribution shifts. First, we propose a data-efficient dual-encoder design with visual foundation models, eliminating the reliance on large-scale scene-text pretraining. Second, we introduce Step-wise LoRA adaptation (SWLoRA), which performs progressive low-rank refinement with a dynamic early-exit mechanism for effective feature adaptation. Third, we propose Common Subspace Fusion (CSF) to align and fuse the two branches in a shared subspace while retaining complementary, shift-robust information. Finally, we construct adverse-condition scene text detection datasets to address the gap in evaluating under imaging degradation. Experiments show that TextDS achieves competitive performance in scene text detection, demonstrating robustness across domains and adverse imaging conditions with only 4.9M trainable parameters.

Visual prompt tuning has emerged as a parameter-efficient fine-tuning approach for adapting large-scale Vision Transformers (ViTs) to downstream tasks. As its learnable prompts are applied in input and feature spaces, prior to jointly going through attention in transformer layers, the most commonly used scheme for fusing image and prompt tokens is concatenation or addition. In this paper, we aim to study a fundamental yet essential problem in visual prompt tuning: whether a single fusion scheme tends to yield better results, and whether that would be beneficial to develop a hybrid fusion scheme. To this end, we formulate the task as a bi-level optimization problem, and solve it leveraging differentiable architecture search. In this context, the learnable prompts and their fusion schemes are jointly optimized. To enrich the search space in the architecture search, we propose two additional fusion schemes, namely, affine transformation and cross-attention, in addition to concatenation and addition. Extensive experiments on 34 datasets spanning VTAB-1k, FGVC, and HTA show consistent gains over prompt-tuning baselines. With a frozen ViT backbone, our method delivers a favorable accuracy--latency--parameter trade-off compared with VPT-Deep and recent variants. Our findings reveal that how prompts fuse with image tokens plays a significant role in visual prompt tuning, and a hybrid fusion fashion can more effectively leverage layer semantics of ViTs, contributing a novel perspective for visual prompt-tuning research.

Vision Foundation Models (VFMs) have significantly advanced dense feature matching, yet severe in-plane rotation remains a critical challenge. Existing solutions face a fundamental dilemma: data-driven methods require inefficient parameter scaling to implicitly learn rotations, whereas strictly equivariant networks lack the semantic capacity of modern VFMs. Consequently, current frameworks typically freeze VFMs and shift the entire burden of rotation generalization to the downstream decoder. To break this architectural bottleneck, we propose REDI-Match, an efficient framework driven by a novel Rotation-Equivariant Distillation (REDI) paradigm. Instead of relying on rotation data augmentation to establish rotational correspondences, REDI distills the non-equivariant semantic representations of a VFM into a lightweight, strictly rotation-equivariant encoder, leveraging an equivariant geometric architecture to constrain robust high-dimensional semantics. To fully exploit these features, we equip the decoder with an entropy-driven spatial alignment module. By evaluating discrete rotation hypotheses, this mechanism explicitly locks onto the canonical coordinate system, eliminating global ambiguity before continuous refinement. Extensive experiments demonstrate that REDI-Match establishes a new state-of-the-art (SOTA) across multiple benchmarks. Notably, it achieves a 13.89% absolute pose accuracy improvement on the highly challenging SatAst dataset while operating 1.9x faster than the current SOTA (RoMa v2), enabling real-time inference (~41 FPS) on a single RTX 4090 GPU. Code: https://github.com/YinjiGe/REDI-Match.

Accurate pulmonary vessel segmentation remains challenging due to the sparse, tortuous, and multi-scale nature of vascular structures, where small branches are easily lost and topology integrity is difficult to preserve under voxel-wise supervision. Existing deep segmentation models primarily optimize binary masks, lacking explicit geometric constraints, thus struggling to recover continuous tubular morphology and fine vascular connectivity. In this study, we introduce MorVess, a morphology-aware segmentation framework that integrates differentiable geometric priors with large-scale foundation model adaptation to achieve fine-grained vascular parsing. MorVess jointly predicts vessel masks, distance maps, and thickness maps, providing explicit supervision for vascular boundaries, centerline consistency, and smooth diameter transitions. A lightweight 2.5D adapter bridges 3D spatial context and 2D SAM representations, while a global-local fusion block aggregates multi-level semantics and geometric cues for high-fidelity topology reconstruction. Across two challenging pulmonary CT benchmarks, MorVess delivers superior Dice, clDice, and HD95 scores, substantially improving small-vessel recovery and global connectivity. These results demonstrate that embedding geometric intelligence into pretrained vision models offers a principled and scalable pathway toward precise vessel analysis and clinically reliable structural quantification. Our source code is available at https://github.com/MaoFuyou/MorVess.

Timely, high-resolution maps of flood extent around settlements are essential for emergency response and damage assessment. We consider airborne RGB imagery for flood mapping as it can be collected rapidly at low cost. To produce flood maps, deep learning models for water segmentation are often used. CNN based and small vision transformer models are used. However, they need much data for adaptation to a change of scenery, i.e., another flooding event. Vision foundation models or large vision transformers are known to generalize across domains. Recently, foundation models for Earth observation became available. They are pretrained on satellite data, whose spatial resolution, viewing geometry, and radiometry differ from nadir RGB imagery. Thus, adaptation is required. We investigate how a satellite-pretrained Earth observation foundation model can be adapted to centimeter-scale floodwater mapping from RGB imagery. Specifically, we fine-tune a model we call Prithvi-2.0-UPN consisting of the Prithvi-EO-2.0-600M Vision Transformer combined with a UPerNet decoder for binary water segmentation on two RGB datasets (BlessemFlood21, NeuenahrFlood). In a first experiment we observe that Prithvi-2.0-UPN reaches state-of-the-art results on BlessemFlood21 and NeuenahrFlood, when trained on their datasets. In a second experiment we show that Prithvi-2.0-UPN performs better than state-of-the-art baseline models for transfer to a new flood event (trained on BlessemFlood21, tested on NeuenahrFlood) in a zero-shot setting. However, the performance indicates room for improvement. In this respect, we investigate in a third experiment how performance improves when further fine-tuning the models with small shares of NeuenahrFlood training data: Prithvi-2.0-UPN improves the fastest and reaches almost the performance level when fully trained on NeuenahrFlood, indicating transfer capabilities.

This report presents our solution for the ICRA 2026 GOOSE 2D Fine-Grained Semantic Segmentation Challenge, which requires parsing unstructured outdoor scenes from four camera platforms into 56 fine-grained categories. Our approach pairs foundation vision encoders (including DINOv3, SigLIP2, and InternImage) with a Mask2Former decoder, and trains them with a strong recipe including long training schedules, exponential moving average, a larger crop size, and multi-scale plus flip test-time augmentation. The three encoders, chosen for their complementary pretraining objectives, are combined into a pretraining-diverse ensemble through per-class validation-IoU weighting. Evaluated on the official GOOSE test set, our submission achieves 75.40% composite mIoU and wins the second place of the challenge. Our study further shows that the encoder's pretraining recipe, rather than its parameter count or the decoder design, is the dominant factor for accuracy on this benchmark.

Concept segmentation models like Segment Anything Model 3 (SAM3) show strong generalization on natural images, yet their performance degrades in medical imaging due to the domain gap caused by different imaging principles and styles. Test-Time Adaptation (TTA) is essential for improving the testing performance by updating the model on the fly without annotations. However, existing vision-language TTA methods are mainly driven by image-level uncertainty minimization, which does not necessarily reflect region-level semantic correctness in medical segmentation. Moreover, they often lack mechanisms to maintain stability in continual one-pass adaptation, leading to limited performance when reliable dense supervision is missing for segmentation. To address these issues, we propose Concept Alignment Contrast and LongShort Prompt Memory for Test-Time Adaptation (CM-TTA) of SAM3 for medical images. First, for a test sample with multiple augmentations, we introduce a novel Concept Alignment Contrast (CAC) metric, which leverages textual-visual semantic consistency to robustly evaluate prediction quality to select the best augmented view as the supervision. Second, to balance rapid and stable adaptation, we design a Long-Short Prompt Memory (LSPM) module. The short memory dynamically fuses recent prompts based on CAC scores for agile local adaptation, while the long memory maintains a stable global prompt to generate enhanced pseudo-labels. Finally, a Densely Supervised Prompt Update (DSPU) strategy is proposed to optimize the prompt embeddings with enhanced pseudo labels as dense supervision. Extensive experiments on prostate and skin lesion segmentation demonstrate that our CM-TTA framework significantly outperforms existing methods for TTA of SAM3. The code is available at https://github.com/SherlockZYB/CM-TTA.