Atif Belal, Lilian Hollard, Marco Pedersoli +1cs.CV
Vision-language object detectors (VLODs) achieve strong zero-shot performance but remain vulnerable to distribution shifts during deployment. Mean-teacher methods for test-time adaptation (TTA) can improve robustness by updating a student model using teacher-generated pseudo-labels. However, mean-teacher TTA is highly sensitive to the choice of a fixed exponential moving average (EMA) coefficient for teacher updates, and repeated optimization with noisy pseudo-labels can cause cumulative student drift. We propose Dynamic EMA and Source Anchoring for TTA (DESA-TTA), a low-overhead method that jointly regulates teacher updates and student drift through dynamic temporal averaging and source anchoring. Dynamic temporal averaging estimates teacher uncertainty from pseudo-label confidence and box density and uses it to select a sample-wise EMA coefficient within bounds determined by teacher parameter drift. Source anchoring partially restores the updated student parameters toward their pretrained values, with the anchoring strength increasing according to student drift. Experiments across diverse distribution shifts and two VLOD architectures show consistent improvements over existing TTA methods. On VOC-C, DESA-TTA improves AP$_{50}$ by 14.5 points over zero-shot inference while achieving 55\% higher inference throughput than the previous state-of-the-art TTA method for YOLO-World. Our code: https://github.com/imatif17/DESA-TTA
Chandler Timm C. Doloriel, Yunbei Zhang, Muhammad Salman Siddiqui +4cs.CV cs.LG
Test-time adaptation (TTA) promises robustness under distribution shift by updating a pretrained model on unlabeled test data, but strict online TTA with batch size one and no access to source data is especially prone to drift or collapse. We introduce Sensitivity-Guided Erasing Adaptation (SEGA), a method for strict online continual TTA (CTTA) on corruption-style streams. SEGA uses a small number of structured erasures to probe how predictive entropy changes as information is removed, and uses the resulting per-sample sensitivity trajectories to coordinate recovery and sample selection rather than relying on raw entropy or batch statistics. This yields a practical feedback signal for long-horizon batch-size-one adaptation without periodic resets or model reservoirs. In experiments on ImageNet-C, CIFAR10/100-C, and corruption-generated aquaculture streams treated as controlled corruption-style proxies, SEGA yields consistent robustness and stability gains over strong CTTA baselines while reducing backward passes through sensitivity-based gating.
We aim to improve frozen DINOv3 dense-prediction models under distribution shift by adding inference computation inside the visual backbone, without changing model weights, task adapters, or prediction heads. The challenge is that repeated transformer-block computation must refine dense features without disrupting the pairwise patch relations that DINOv3 uses to preserve spatial structure. We introduce GramLoop, a training-free framework that replays a short transformer window and controls each replay through final-layer cosine-Gram consistency. Each proposal is propagated through the frozen suffix, measured against the standard DINOv3 trajectory, and accepted through a patchwise gate at the replay-window endpoint. Across object detection and semantic segmentation under corruptions, perturbations, and natural shifts, GramLoop improves all five shifted benchmarks over the paired DINOv3 baseline. On COCO-O, it improves mAP by +0.252 and Effective Robustness by +0.250, while preserving clean ADE20K performance. Code will be released at https://github.com/cheyan9/GramLoop.
Recent advances in anomaly detection (AD) for industrial inspection have pushed performance on standard benchmarks toward saturation. However, strong benchmark performance does not necessarily translate to real-world deployment, as these benchmarks are primarily collected under controlled acquisition conditions. Changes in illumination, background, viewpoint, and other environmental factors can shift normal samples away from the learned normal distribution and cause false anomaly responses. We address AD under such distribution shifts by explicitly modeling nuisance variation from changing imaging conditions in feature space. Without anomaly labels or target-domain data, our Nuisance-Filtered Anomaly Detection (NFAD) framework estimates a nuisance subspace from matched feature displacements induced by content-preserving perturbations and suppresses its contribution to anomaly residuals at inference. The same subspace supports two complementary branches: full projection for image-level detection and selective suppression for pixel-level localization, preserving evidence of localized defects. On AeBAD-S, a benchmark specifically designed for AD under acquisition shifts, NFAD achieves 91.0\% image-level AUROC, establishing a new state of the art. Notably, this robustness does not come at the expense of conventional AD performance: NFAD remains competitive on standard benchmarks that do not explicitly evaluate distribution shift, including VisA, Real-IAD, and MVTec AD. These results show that explicitly suppressing such nuisance variation improves AD under distribution shift while preserving strong performance in standard settings.
Test-time adaptation (TTA) has been widely explored in single-label recognition, effectively mitigating distribution shifts, especially when combined with vision-language models. However, real-world images often contain multiple objects, while the more practical multi-label test-time adaptation (MLTTA) has received little attention so far. Recent cache-based TTA methods have shown promising efficiency and effectiveness, yet directly extending them to multi-label scenarios suffers from a one-to-many mapping problem: a shared global representation entangling co-occurring objects is stored as class-wise cache prototypes, inducing dominant-label bias and compromised cache calibration. While introducing region-level cues helps isolate class-specific evidence, such regional evidence can also be unreliable under distribution shifts, making its identification and utilization non-trivial. To address these issues, we introduce PuRF, a novel PuRiFication-driven cache-based method for multi-label test-time adaptation of vision-language models. Specifically, PuRF first performs region purification to identify reliable regions, providing comprehensive regional cues for multi-label recognition and enabling fine-grained alignment. Based on these purified regions, PuRF conducts cache purification to enhance cache representation and adaptability, where episodic purification builds a discriminative region-based cache, and temporal refreshing further promotes long-term cache adaptability. Experiments demonstrate that PuRF consistently outperforms state-of-the-art methods, achieving a notable 4.05% mAP improvement on ViT-B/32 across five datasets.
Test-time adaptation (TTA) aims to improve model robustness under distribution shift by adapting a source model using unlabeled test data. Although methods such as TENT and EATA have demonstrated gains on corrupted data, aggregate accuracy can obscure the conditions under which adaptation fails or provides little benefit. We present a controlled comparison of three TTA strategies---BatchNorm-statistics adaptation (BN-Adapt), entropy-minimization adaptation (TENT), and reliability-filtered adaptation (a scoped re-implementation of EATA)---against an unadapted source model on the full CIFAR-10-C benchmark, covering 15 corruption types and 5 severity levels. All three methods improve mean accuracy over the source model by 12.2--13.3 percentage points (Wilcoxon signed-rank $p < 10^{-12}$). However, each method underperforms the source model on 8.0--9.3\% of conditions, with failures concentrated in low-severity corruptions where the source model already performs near ceiling, particularly brightness, fog, contrast, and defocus blur. We further find that EATA closely tracks the gradient-free BN-Adapt baseline, with a mean absolute difference of 0.09 percentage points, compared with 1.08 percentage points relative to TENT. This suggests that reliability filtering can substantially restrict effective adaptation, causing EATA to behave more like a BatchNorm-statistics baseline than an entropy-minimization method. These results show that aggregate accuracy alone can mask systematic TTA failure modes and motivate condition-level evaluation of when adaptation helps, harms, or becomes effectively inactive.
Vision-based industrial anomaly detectors are calibrated on one distribution but may be deployed on another that differs in illumination, fixture placement, or sensor characteristics, sharply degrading an otherwise accurate detector. Adapting to the incoming lot is a natural response, but labeled anomalies are scarce. We therefore consider calibration using only a handful of verified-normal images available before scoring the rest of the lot. Existing fixes require backpropagation, detector-specific tuning, or choices about feature directions that few calibration samples cannot justify. We present SPARC, a few-shot calibration method that intercepts patch features between encoder and detector and removes a closed-form, spatially indexed estimate of deployment-time nuisance through per-cell subspace projection. It needs only $k \le 8$ verified-normal images and uses the algebraic saturation rank $r{=}k{-}1$ on the encoder's native patch grid. The correction requires no gradient or weight updates and works with memory-bank, density, prototype, and mutual detectors. On the shift-prone benchmarks, SPARC improves pooled Image AUROC and AU-PRO$_{0.3}$ for all seven detectors whose image scores depend on corrected patch features by $+13.8$ and $+3.5$ percentage points (pp), respectively; on benchmarks without engineered shift, the changes are small and mixed. Controls that give competing corrections the same calibration images attribute these gains to the per-cell subspace structure rather than the images alone. Further ablations support the saturation-rank choice and characterize sensitivity to backbone and calibration conditions.
Prediction cascades significantly reduce energy consumption of Artificial Intelligence (AI) models while maintaining high predictive performance. The idea is that easy inputs are routed through a lightweight small model, and difficult uncertain cases are deferred to a larger model. While this design can improve computational efficiency on clean data, its effectiveness depends on the reliability of confidence-based routing. Input degradations, such as static corruptions and sequential perturbations, can shift model confidence and routing decisions. In this paper, we study confidence-based cascade frameworks for image classification and investigate how such degradations affect their confidence-based deferral behavior. We select a model cascade at the pareto-optimum of accuracy, routing quality, and energy consumption that achieves competitive predictive performance with an up to 10-fold decrease in CO$_2$ emissions. We study the behavior of that model cascade under input corruptions and analyze how the cascade's routing decisions change when the input distribution shifts. Our analysis identifies three failure modes. Static corruptions either (1) break the routing signal while the large model remains useful, or (2) degrade both models so deferral no longer recovers accuracy. Sequential perturbations reveal a third mode: predictions stabilize but deferral suppresses, yielding stable but unreliable predictions. These findings demonstrate that energy efficient model cascades require evaluation beyond clean accuracy, with explicit attention to routing reliability under distribution shift.
Nils Lehmann, Jakob Gawlikowski, Burak Ekim +2cs.CV
Geospatial Foundation Models (GeoFMs) are most commonly ranked and selected by accuracy on standard benchmark conditions via averaged ranks. We show that this protocol is too narrow: the promised deployment in critical EO tasks requires further angles of analysis, mainly calibration, the agreement between a model's confidence and its correctness. Across 16 frozen encoders, four classification and five segmentation datasets, and two orthogonal stress axes, every encoder degrades as corruption intensifies, and the ranking changes as well. Across the four classification benchmarks, EO-pretrained and ImageNet-pretrained encoders are indistinguishable on clean accuracy and clean calibration, and EO pretraining provides no more stability under shift than ImageNet pretraining. Under shift the GeoFMs drift further into overconfidence than the ImageNet-pretrained encoders, at every grade and in every corruption family. A centered kernel alignment (CKA) analysis ties this to representational rigidity: EO-pretrained embeddings move less under corruption while losing just as much task information and remaining overconfident. We apply three commonly explored uncertainty quantification methods and find that temperature scaling and deep ensembles cannot counteract the degradation, while a Gaussian-process probe roughly halves ECE under severe cloud only by tripling it on clean data. In selective prediction experiments, we find that confidence-based abstention cannot defer around confidently wrong predictions, and advocate that benchmark rankings and evaluations should therefore operate across a multitude of conditions and metrics to more holistically evaluate model development progress and close the gap to real world deployment scenarios.
Test-time adaptation (TTA) can improve the recognition accuracy of vision-language models under distribution shift, but often degrades calibration, making predictive confidence unreliable for downstream decision-making. Many existing label-free calibration approaches are either coupled to prompt optimization or rely on logit-range statistics that provide only a coarse characterization of the predictive distribution. We show that TTA can increase confidence and reduce entropy even when the top-1 prediction and its correctness remain unchanged, a failure mode we term prediction-preserving sharpening. Across diverse TTA methods and benchmarks, larger entropy reductions relative to paired zero-shot predictions are associated with greater increases in Expected Calibration Error (ECE). On entropy-reduced samples, confidence gains also tend to exceed accuracy gains. Based on these findings, we propose Zero-Shot-Anchored Entropy Calibration (ZAEC), a label-free post-hoc method that uses zero-shot entropy as a sample-specific uncertainty reference. ZAEC selectively restores the zero-shot entropy of sharpened predictions through minimal temperature scaling while leaving all other predictions unchanged. It requires no labeled calibration data or learned parameters and preserves class rankings and classification accuracy. Across five TTA methods and 15 datasets, ZAEC achieves the lowest post-hoc macro-average ECE on ViT-B/16, with consistent gains on RN50.
Vladan Stojnić, Ryan Ramos, Giorgos Kordopatis-Zilos +2cs.CV cs.LG
Deep vision models exploit shortcuts, relying on cues that correlate with supervision signals. Prior work has focused on visible biases, such as object-background or texture correlations. We identify a different source of shortcut learning: invisible metadata traces embedded at the pixel level, for metadata such as image processing and photo acquisition. We hypothesize that large-scale semantic supervision, whether through categorical labels (ImageNet) or billion-scale captions (LAION), naturally induces metadata-semantics correlations during pretraining, leading models to convert low-level signals into predictive features. By introducing controlled metadata-semantics correlations, we show that stronger ones produce systematically higher sensitivity to metadata traces and larger performance degradation under metadata distribution shifts. We further explore mitigation strategies applied during and after pretraining that reduce sensitivity not only to targeted metadata but also to unseen ones, without sacrificing performance on downstream tasks. Metadata sensitivity also has a positive side: it partly explains the strong generated-image detection ability of some encoders, while its mitigation can improve out-of-distribution generalization. Code: https://github.com/ryan-caesar-ramos/visual-encoder-traces
Malena Loza, Felipe Grijalva, Eva Milara +3cs.CV cs.LG
Tabular-to-image methods that convert tabular data into visual representations have emerged as a novel paradigm for leveraging the high performance of deep learning models. Despite their advantages, the robustness of these methods under distribution shifts remains under explored. Test-Time Augmentation (TTA) is an effective approach in image classification to improve model generalization and robustness, where predictions over multiple transformed views of each input are aggregated. This work evaluates the impact of TTA techniques on predictive performance under Out-Of-Distribution (OOD) for representations generated by tabular-to-image methods. Six tabular-to-image encoding methods were considered: TINTO, IGTD, DeepInsight, BIE, DistanceMatrix, Fotomics. Twenty-five TTA techniques were used, organized into six types: Geometric, Photometric, Structural, Frequency/Encoding, Mixup, and Composite. We employed two datasets from the TableShift benchmark (HELOC and Voting) that provide in-distribution and OOD test subsets designed to evaluate the effect of distribution shifts on tabular data. The results indicate that TTA improves OOD performance, with composite and photometric strategies providing the best trade-off between robustness and variance. In contrast, frequency-domain transformations that alter the encoder's feature-to-intensity mapping consistently degrade performance. These findings highlight TTA as a promising approach for improving the robustness and generalization of classifiers trained on image representations derived from tabular data, particularly under distribution shifts.
Vision-language models (VLMs) such as CLIP exhibit remarkable zero-shot capabilities, yet their performance frequently degrades sharply under unexpected test-time distribution shifts. While Test-Time Adaptation (TTA) offers a promising solution, continuously adapting VLMs over an unlabeled test stream presents fundamental challenges. Conventional top-1-centric updates often reinforce errors by corrupting the local semantic geometry among related classes, while iterative adaptation exacerbates progressive bias accumulation, ultimately driving the model toward mode collapse. To overcome these coupled vulnerabilities, we propose Local Margin Restoration (LMR), a lightweight, one-step TTA framework. At the sample level, our Protected Margin Restoration (PMR) objective recovers local semantic geometry by shielding plausible near-top candidates from external hard negatives. Concurrently, to combat stream-level degradation, we introduce a dual-stage stabilization mechanism, featuring an Adaptive Margin (AM) controller and Bias Correction (BC), to dynamically disrupt progressive bias accumulation and prevent mode collapse. Extensive experiments on CIFAR-C, ImageNet-C, and ImageNet variants demonstrate that LMR consistently outperforms state-of-the-art TTA baselines, proving exceptionally robust and efficient even in challenging low-batch test-time regimes. Our code is available at https://github.com/DennisHuangYan/LMR.
A popular route to interpretable zero-shot classification asks a large language model (LLM) to describe each class name and prompts CLIP with the resulting descriptors. We show that these descriptors carry little visual evidence of their own: removing the class name from the prompt collapses ImageNet accuracy from 59.5% to 15.5%. The diagnosis is that the descriptors are conditioned on the label rather than on the images, so they describe the concept in general and mislead exactly when the data shifts; an LLM insists that strawberries are red, but every strawberry in ImageNet-Sketch is a colorless line drawing. We therefore select attributes from the target image collection instead: we score a large attribute pool against the images in CLIP's joint embedding space and keep the top-scoring attributes per class. Selected this way, class-name-free attribute prompts reach 23.8% on ImageNet (against 15.5% for LLM descriptors), the gain holds on four shifted ImageNet variants, and reselecting from the LLM's own pool isolates the selection mechanism as the cause. With one image per class, the selected attributes outperform the prompt-tuning method CoOp by 3 points while fitting in under a minute instead of 14 hours, with no learned soft prompt to obscure the decision. Because the attribute set is chosen by the data, it doubles as a readable summary of a dataset, which we use to describe distribution shift in words.
Deep neural networks trained with Empirical Risk Minimization (ERM) often fail under distribution shifts because they exploit spurious correlations between object labels and background context. Recent generative approaches address this issue by creating counterfactual images with altered contexts, but typically use these samples as standard data augmentation, leaving the model free to retain background-sensitive representations. We propose a two-stage framework that uses generative intervention to explicitly learn background-invariant visual representations. First, we isolate the foreground object using zero-shot segmentation and generate context-shifted variants with a structure-preserving diffusion model, preserving object identity while varying the surrounding environment. We then introduce Cross-Variant Self-Supervised Learning, where variants of the same object under different backgrounds form positive pairs in a contrastive objective. This encourages the encoder to align object-centric representations while suppressing background-specific cues. Then, we fine-tune the pretrained encoder using an ERM warm-up followed by GroupDRO with layer-wise learning rates. Experiments on distribution-shift benchmarks demonstrate best worst-group performance, achieving 92.5% on Waterbirds, 81.7% on MetaShift, and 87.4% on NICO++. Code: https://github.com/surajyadav-research/GRSSL
Test-Time Adaptation (TTA) seeks to improve model robustness under distribution shifts by adapting parameters using unlabeled target data. However, in the absence of supervision, entropy-based adaptation is fundamentally underconstrained: multiple distinct parameter updates can achieve similarly low entropy while inducing drastically different decision boundaries. This phenomenon, known as underspecification, renders standard TTA brittle and prone to collapse into spurious modes. In this work, we reinterpret TTA through a posterior-inspired lens induced by entropy minimization, where low-entropy solutions define a pseudo-likelihood over parameters. Instead of committing to a single point estimate, we introduce a particle-based diversification framework that explores multiple plausible adaptation trajectories simultaneously. Our method can be viewed as a structured exploration of multiple plausible adaptation solutions, implemented through multi-level diversification at the output, parameter, optimizer, and input levels. Crucially, the framework acts as a plug-and-play wrapper compatible with existing TTA methods. Extensive experiments on challenging benchmarks demonstrate consistent gains in stability and robustness, achieving improvements of 3-4% under mixed shifts, 2-3% with batch size one, and 1-2.5% under label shifts, outperforming state-of-the-art baselines. Our results suggest that treating TTA as a multi-hypothesis inference problem, rather than a single-point optimization task, is key to mitigating underspecification and enabling reliable real-world deployment.
Zhuocheng Shang, Sanmay Das, Ahmed Eldawycs.CV cs.LG
Geospatial foundation models pretrained on satellite imagery promise broad generalization across remote sensing tasks and regions, but their geographic transferability has not been systematically tested, especially in agriculture applications. This paper presents a controlled benchmark that evaluates three models, Prithvi, SpectralGPT, and SatMAE, on multi-temporal crop segmentation and change detection across four U.S. states, Iowa, North Carolina, California, and Minnesota. By assigning each train, validation, and test split to a separate region, we measure how well each model transfers to land it has not seen. All three degrade sharply under regional distribution shift, predicting only the most common crops while missing rare ones. We further find that fitting these models to a shared input format affects each one differently, which complicates direct architectural comparison. These results expose key limitations of current geospatial foundation models for agriculture and point to region aware evaluation as a necessary standard.
Test-Time Adaptation (TTA) methods commonly update the affine parameters of normalization layers to adapt deployed models under distribution shifts. However, per-channel affine parameters perform axis-aligned scaling and shifting, making them geometrically incapable of correcting cross-channel structural changes induced by distribution shift. To address this limitation, we propose MixTTA, a lightweight plug-in module that equips normalization layers with a low-rank cross-channel transformation, enabling inter-channel mixing at each layer. To ensure that the low-rank branch captures only cross-channel interactions, we also propose Decoupling Projection that enforces strict separation from the diagonal affine path, along with Spectral Projection that prevents rank-1 collapse under non-stationary test streams. MixTTA can be seamlessly integrated into any existing normalization-based TTA method. Experiments in both standard and wild TTA settings show consistent improvements over strong baselines while mitigating adaptation failure under challenging conditions. The source code is publicly available at https://github.com/delta6189/MixTTA.
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.
Facial expression recognition (FER) is inherently ambiguous: human annotators frequently disagree, and models deployed in real environments face distribution shift. Crucially, these two conditions demand different downstream actions, as ambiguous in-distribution faces should be reported with their ambiguity whereas out-of-distribution inputs should be rejected. However, a single uncertainty score conflates the two. In this study, uncertainty decomposition into aleatoric and epistemic components for FER is investigated, and Uncertainty-Aware Routing (UAR), an inference-time routing mechanism that exploits the separation, is introduced. Specifically, aleatoric and epistemic uncertainties are obtained from a Deep Ensemble of fully fine-tuned DINOv2 models and are each validated against an independent external signal: aleatoric against human annotator disagreement, and epistemic against distribution shift induced by image corruptions. The proposed dual-validation protocol reveals that aleatoric recovers annotator disagreement with Spearman correlation 0.66 (95% CI: 0.64-0.68), and epistemic detects corruption-induced shifts, achieving average AUROC of 0.699 at the highest corruption severity. UAR retains approximately 1.8 times more ambiguous in-distribution faces than single-uncertainty routing at a matched out-of-distribution rejection rate. A strong label-distribution-learning baseline achieves comparable disagreement recovery but cannot separate ambiguity from shift and therefore cannot route, establishing that the value of decomposition lies in the separation enabling interpretable and differentiated action selection.
Recent vision token pruning methods effectively preserve model performance under moderate token budgets but become unstable under ultra-low token budget. Our analysis shows that as the pruning budget decreases, accuracy degradation is often accompanied by larger feature distribution shifts. Critically, the degree of this distribution shift strongly correlates with performance degradation. To better characterize this phenomenon, we introduce a lightweight distribution consistency metric to estimate the distribution shift between retained and full tokens. Motivated by these observations, we propose a two-stage pruning framework consisting of Anchor-Context Graph Recovery (ACGR) and Text-Aware Token Cluster Selection (TATCS). Specifically, ACGR transfers contextual information before token removal, while TATCS dynamically re-selects representative tokens when severe distribution shift is detected. Extensive experiments demonstrate that our method achieves superior and more stable performance under ultra-low token budget. Notably, it retains 92.1% of the upper-bound average performance on LLaVA-1.5-7B with only 16 visual tokens.
Jiazhen Huang, Xiao Chen, Zhiming Liu +3cs.CV cs.LG
Vision-Language Models (VLMs) such as CLIP have become a standard backbone for open-vocabulary recognition, yet their zero-shot predictions remain vulnerable to distribution shifts encountered at deployment. Test-Time Adaptation (TTA) has recently been extended to CLIP as a lightweight solution, leading to a rapidly growing body of TTA4CLIP methods. However, empirical progress in this area has largely outpaced our understanding of what truly drives adaptation, where their gains originate, and under which shifts they remain reliable. In this paper, we take a step back from the pursuit of state-of-the-art accuracy and conduct a systematic controlled study of TTA4CLIP. We first organize existing methods into three unified paradigms according to what is updated at test time. We then introduce TTABC, an open-source TTA Benchmark for CLIP, which standardizes evaluation protocols and integrates more than 20 representative methods. Our controlled empirical analysis focuses on three key areas. First, we determine the driving factors in parameter-based methods, revealing that adaptation gains are primarily driven by test-time evidence and reliable proxies rather than heavy optimization. Second, we explore evidence utilization beyond heavy parameter tuning, showing that competitive and efficient performance can be achieved through cross- or current-sample evidence and lightweight prototype updates. Finally, we demonstrate that there is no silver bullet for TTA: no single adaptation paradigm is universally optimal, and the preferred paradigm depends on the nature of shift. We hope our benchmark and study provide a clearer understanding of the current TTA4CLIP landscape and establish a foundation for further research.
Multi-label recognition with frozen Vision-Language Models (VLMs) is brittle under distribution shift: standard zero-shot inference scores labels independently, ignoring co-occurrence structure and producing incoherent label sets where dominant concepts suppress weaker but compatible labels. We introduce Bayesian Conditional Priors (BCP) Estimation, a gradient-free test-time adaptation method that injects label dependency without tuning the backbone. BCP views zero-shot logits as a proxy for marginal posteriors under a fixed image-text likelihood and attributes shift-induced errors mainly to a mismatched label prior. For each test image, it selects a high-confidence anchor label and applies an anchor-conditioned Bayesian refinement. This update is closed-form in logit space and admits a pointwise mutual information (PMI) interpretation, explicitly promoting compatible labels and suppressing incompatible ones. BCP operates without target annotations by estimating anchor-conditioned priors online from the unlabeled test stream via lightweight second-order co-occurrence statistics, adding negligible overhead beyond a single forward pass. Across standard multi-label benchmarks and multiple CLIP backbones, BCP consistently outperforms strong TTA baselines, e.g., improving RN50 average mAP from 57.31 to 69.22 and ViT-B/16 from 62.61 to 71.79.
Generated (or synthetic) image data is increasingly used to augment or replace real training datasets when target imagery is scarce, expensive, or biased. For hand detection, particularly in occupational safety settings, public datasets mostly contain bare hands. This under-represents the variation in hand appearance introduced by gloves, tattoos, jewelry, and other personal protective equipment, creating a distribution shift that safety-critical applications encounter at deployment. We test whether generative inpainting, editing only the hand region of a real photograph to introduce accessories, can close this shift gap. On a paired dataset of real images and their synthetic counterparts, we train YOLOv8n hand detectors under six training-and-scheduling regimes (Experiments A-F, three random seeds each), evaluate every detector on a real test set and on a real-gloves-only test split, and report the mean average precision (mAP) at two overlap thresholds (mAP@0.5 and mAP@0.5:0.95) along with paired statistical tests. A two-stage experiment: train on real U synthetic data, then fine-tune the resulting weights on real-only at a lower learning rate, increases mAP@0.5 compared to the real-only baseline model on the standard real test set, and improves the real-gloves out-of-distribution gap. Another three-stage experiment preserves box-tightness best, reaching the highest mAP@0.5:0.95 of any other experiment in the study. The synthetic-data utility for safety-critical hand detection is determined by the training procedure, and simple multi-stage experiments extract substantial real-deployment benefit from inpainted accessory data.