Robust visual recognition on embedded platforms requires models that both generalize out-of-distribution (OOD) and fit into tiny compute/memory budgets. While pre-training is a standard route to robustness for mid/large backbones, its value in the ultra-small regime remains unclear. We present a capacity-aware study of pre-training for two efficient ConvNet families (EfficientNet and MobileNetV3) scaled from "small" to "ultra-small" via a simple, reproducible recipe. We compare three initializations — ImageNet→COCO pretraining, ImageNet classification pretraining, and training from scratch—on two axes of distribution shift: (i) cross-dataset RGB→RGB transfer between LLVIP and FLIR (ii) cross-modality detection where models are fine-tuned on RGB and evaluated on infrared (IR). A complementary classification study on DomainNet probes whether the trends extend beyond detection. Across settings, we find that pretraining’s benefit is conditional on both backbone capacity and shift difficulty. Task-aligned Imagenet→COCO pretraining is the most reliable starting point at moderate sizes and for the easier transfer direction. In the low-capacity regimes, differences are typically within run-to-run variation, and training from scratch can match or surpass pre-training. Classification mirrors this capacity gating. Our results test the premise "pretraining always helps" and instead quantify when task-aligned pretraining pays off for ultra-small backbones and when it likely does not 1.

Personalized expression recognition (ER) involves adapting a machine learning model to subject-specific data for improved recognition of expressions with considerable inter-personal variability. Subject-specific ER can benefit significantly from multi-source domain adaptation (MSDA) methods – where each domain corresponds to a specific subject – to improve model accuracy and robustness. Despite promising results, state-of-the-art MSDA approaches often overlook multimodal information or blend sources into a single domain, limiting subject diversity and failing to explicitly capture unique subject-specific characteristics. To address these limitations, we introduce MuSACo, a multi-modal subject-specific selection and adaptation method for ER based on co-training. It leverages complementary information across multiple modalities and multiple source domains for subject-specific adaptation. This makes MuSACo particularly relevant for affective computing applications in digital health, such as patient-specific assessment for stress or pain, where subject-level nuances are crucial. MuSACo selects source subjects relevant to the target and generates pseudo-labels using the dominant modality for class-aware learning, in conjunction with a class-agnostic loss to learn from less confident target samples. Finally, source features from each modality are aligned, while only confident target features are combined. Experimental results on challenging multimodal ER datasets – BioVid, StressID, and BAH – show that MuSACo outperforms UDA (blending) and state-of-the-art MSDA methods. Our code is available: https://github.com/osamazeeshan/MuSACo

Facial expression recognition (FER) in videos is challenging because models must identify subtle, temporally evolving affective states that vary across individuals. Although vision-language models provide transferable visual-semantic representations, models trained on subject-independent data often degrade under subject-specific distribution shifts at inference time. Existing test-time adaptation (TTA) methods commonly update model parameters during inference, increasing computational cost and latency. Cache-based methods avoid parameter updates, but they usually require enough target samples to form reliable class prototypes, which is difficult early in adaptation and for rarely observed classes. We introduce Energy-Based Cache Personalization (EB-CaP), a subject-based online TTA method for video FER that generates class-specific prototypes personalized to each target video. EB-CaP uses a lightweight energy-based model to sample prototypes from the current unlabeled video and populate a personalized cache online, without accumulating large amounts of target data or storing diverse source prototypes. Its energy function relies only on pretrained CLIP: similarities between the target video embedding and class text embeddings guide prototype sampling. In parallel, positive and negative caches store reliable and uncertain target embeddings. An adaptive entropy gate controls cache updates according to the evolving confidence distribution, while a diversity gate limits redundant samples. Final predictions combine cache-derived scores with the current CLIP scores. Experiments on BioVid, StressID, and BAH show that EB-CaP outperforms state-of-the-art TTA methods while maintaining low computational and memory overhead. Code is available at https://github.com/MasoumehSharafi/EB-CaP.

Multimodal learning has shown promise in medical imaging, combining complementary modalities like images and text. Vision-language models (VLMs) capture rich diagnostic cues but often require large paired datasets and promptor text-based inference. Their practicality is therefore limited due to annotation cost, privacy, and compute demands. Unpaired external text, like pathology reports, can still provide complementary diagnostic cues if semantically relevant content is retrievable per image. To address this, we introduce CLIP-IT, a novel framework that relies on rich unpaired text reports. Specifically, CLIP-IT uses a CLIP model pre-trained on histology image–text pairs from a separate dataset to retrieve the most relevant unpaired textual report for each image in the downstream unimodal dataset. These reports, sourced from the same disease domain and tissue type, form pseudo-pairs that reflect shared clinical semantics rather than exact alignment. Knowledge from these texts is distilled into the vision model during training, while LoRA-based adaptation mitigates the semantic gap between unaligned modalities. At inference, only the vision model is used, maintaining low overhead while still benefiting from multimodal training without requiring paired data in the downstream dataset. Experiments1 show that CLIP-IT consistently improves classification accuracy over both unimodal and multimodal CLIP-based baselines in most cases, without requiring paired annotations per dataset or incurring additional inference-time complexity.

Lightstage facial capture produces production-quality digital humans, but it is resource and labor-intensive. Multi-camera setups, hours of computation, and massive data storage create bottlenecks that hinder iterative workflows. This paper introduces FaceSnap, an end-to-end framework that streamlines capture via a two-stage approach. First, a one-time multi-view optimization from a range-of-motion sequence builds a personalized model encoding both geometry and expression-dependent appearance. This model then enables high-fidelity real-time facial performance capture from a single monocular lightstage camera, with no further multi-view capture required. FaceSnap jointly estimates geometry and dynamic 4K texture at 83 fps. The 4K texture is produced by a novel personalized residual upscaler that recovers subject-specific high-frequency detail, which generic upscalers fail to capture. FaceSnap achieves geometric accuracy competitive with full per-frame multi-view optimization while outperforming feed-forward methods trained on production-quality 3D data, all from a single camera view. Finally, we introduce Multi4D, a public benchmark for evaluating 4D facial reconstruction methods in lightstage environments, enabling topology-invariant geometric comparison across methods.

Text-to-image person re-identification (TI-ReID) relies on natural-language text descriptions to retrieve top matching individuals from a gallery of reference images. While recent large vision-language models (VLMs) achieve strong retrieval performance, their decisions remain largely uninterpretable. Existing interpretability approaches in TI-ReID rely solely on slot-attention to highlight attended regions, but fail to reliably bind visual regions to semantically meaningful concepts, limiting interpretation to qualitative visualizations over a restricted vocabulary. This paper introduces InterPartAbility, an interpretable TI-ReID method that performs explicit part-wise matching and enables phrase-region grounding. Unlike parameter-heavy slot-attention methods that yield only qualitative interpretability, our open-vocabulary patch-phrase interaction module (PPIM) guides a standard TI-ReID model with concept-level phrases. Concept-based part phrases provide evidence that encourages the model to attend to the corresponding local image regions. InterPartAbility further leverages CLIP ViT self-attention to produce spatially concentrated patch activations aligned with each part-level phrase, yielding grounded explanation maps. Finally, a quantitative interpretability protocol for TI-ReID is introduced that extends current perturbation-based evaluation metrics into the TI-Reid domain. This includes a counterfactual region removal that measures retrieval degradation when top-ranked explanatory regions are removed. Empirical results on three challenging benchmarks show that InterPartAbility can achieve SOTA interpretability performance under these metrics, while sustaining competitive retrieval accuracy.

Facial expression recognition (FER) models are employed in many video-based affective computing applications, such as human-computer interaction and healthcare monitoring. However, deep FER models often struggle with subtle expressions and high inter-subject variability, limiting their performance in real-world applications. To improve their performance, source-free domain adaptation (SFDA) methods have been proposed to personalize a pretrained source model using only unlabeled target domain data, thereby avoiding data privacy, storage, and transmission constraints. This paper addresses a challenging scenario, where source data is unavailable for adaptation, and only unlabeled target data consisting solely of neutral expressions is available. SFDA methods are not typically designed to adapt using target data from only a single class. Further, using models to generate facial images with non-neutral expressions can be unstable and computationally intensive. In this paper, personalized feature translation (PFT) is proposed for SFDA. Unlike current image translation methods for SFDA, our lightweight method operates in the latent space. We first pre-train the translator on the source domain data to transform the subject-specific style features from one source subject into another. Expression information is preserved by optimizing a combination of expression consistency and style-aware objectives. Then, the translator is adapted on neutral target data, without using source data or image synthesis. By translating in the latent space, PFT avoids the complexity and noise of face expression generation, producing discriminative embeddings optimized for classification. Using PFT eliminates the need for image synthesis, reduces computational overhead (using a lightweight translator), and only adapts part of the model, making the method efficient compared to image-based translation. Extensive experiments on four challenging video FER benchmark datasets, BioVid, StressID, BAH, and AffWild2, show that PFT consistently outperforms state-of-the-art SFDA methods, providing a cost-effective approach that is suitable for real-world, privacy-sensitive FER applications.

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

As AI systems become more capable, it is important that their decisions are understandable and aligned with human expectations. A key challenge is the lack of interpretability in deep models. Existing methods such as GradCAM generate heatmaps but provide limited conceptual insight, while prototype-based approaches offer example-based explanations but often rely on rigid region selection and lack semantic consistency. To address these limitations, we propose PCMNet, a Part-Prototypical Concept Mining Network that learns human-comprehensible prototypes from meaningful regions without extra supervision. By clustering these into concept groups and extracting concept activation vectors, PCMNet provides structured, concept-level explanations and enhances robustness under occlusion and adversarial conditions, which are both critical for building reliable and aligned AI systems. Experiments across multiple benchmarks show that PCMNet outperforms state-of-the-art methods in interpretability, stability, and robustness. This work contributes to AI alignment by enhancing transparency, controllability, and trustworthiness in modern AI systems.

Large video-language models (LVLMs) have shown remarkable performance on multimodal tasks like multimodal emotion recognition (ER) in the wild. ER is inherently multimodal, requiring a joint understanding of facial expressions, vocalizations, language, biosignals, and gestures. However, real-world deployment remains challenging: modalities may be missing or noisy at test time. Partial observations can be viewed as a distribution shift relative to the complete-modality distribution. SOTA TTA methods based on entropy minimization or perplexity reduction do not transfer to autoregressive LVLMs, while retrieval augmented generation (RAG) degrades when the observed modality is weak. Because no ground-truth supervision exists to verify individual updates, adaptation across this stream risks accumulating drift and degrading once the model departs from a reliable solution. An effective solution must therefore adapt to arbitrary missing-modality patterns and remain effective during continual adaptation. We address both jointly with Test-Time Self-Distillation (TTSD), a parameter-efficient framework in which a frozen teacher, trained on complete modalities, guides an adaptive low-rank student via self-distillation, updating only a negligible number of parameters. Stability is built into this same loop through Fisher-Anchored Restoration (FAR), which monitors Fisher information stability to detect convergence versus drift and restores the student toward the teacher's anchor when distributional shifts are identified. Our experiments on MELD, DFEW, and BAH under 0

Personalization in emotion recognition (ER) is essential for accurate interpretation of subtle and subject-specific expressive patterns. Recent advances in vision-language models (VLMs), such as CLIP, demonstrate strong potential for leveraging joint image-text representations in ER. However, existing CLIP-based methods either rely on CLIP's contrastive pretraining or use LLMs to generate descriptive text prompts, which can be noisy, computationally expensive, and often fail to capture fine-grained expressions, leading to degraded performance. In this work, Action Units (AUs) are leveraged as structured textual prompts within CLIP to model fine-grained facial expressions. AUs encode the subtle muscle activations underlying expressions, providing localized and interpretable semantic cues for more robust facial expression recognition (FER). We introduce CLIP-AU, a lightweight AU-guided temporal learning method that integrates interpretable AU semantics into CLIP. It learns generic, subject-agnostic representations by aligning AU prompts with facial dynamics, enabling fine-grained FER without CLIP fine-tuning or LLM-generated text supervision. Although CLIP-AU models fine-grained AU semantics, it does not adapt to subject-specific variability in subtle expressions. To address this limitation, we propose CLIP-AUTT, a video-based test-time personalization method that dynamically adapts AU prompts to videos from unseen subjects. By combining entropy-guided temporal window selection with prompt tuning, CLIP-AUTT enables subject-specific adaptation while preserving temporal consistency. Our experiments on three challenging video-based datasets, BioVid, StressID, and BAH, indicate that CLIP-AU and CLIP-AUTT outperform state-of-the-art CLIP-based FER and TTA methods.

Using behavioural science, health interventions focus on behaviour change by providing a framework to help patients acquire and maintain healthy habits that improve medical outcomes. In-person interventions are costly and difficult to scale, especially in resource-limited regions. Digital health interventions offer a cost-effective approach, potentially supporting independent living and self-management. Automating such interventions, especially through machine learning, has gained considerable attention recently. Ambivalence and hesitancy (A/H) play a primary role for individuals to delay, avoid, or abandon health interventions. A/H are subtle and conflicting emotions that place a person in a state between positive and negative evaluations of a behaviour, or between acceptance and refusal to engage in it. They manifest as affective inconsistency across modalities or within a modality, such as language, facial, vocal expressions, and body language. While experts can be trained to recognize A/H, integrating them into digital health interventions is costly and less effective. Automatic A/H recognition is therefore critical for the personalization and cost-effectiveness of digital health interventions. Here, we explore the application of deep learning models for A/H recognition in videos, a multi-modal task by nature. In particular, this paper covers three learning setups: supervised learning, unsupervised domain adaptation for personalization, and zero-shot inference via large language models (LLMs). Our experiments are conducted on the unique and recently published BAH video dataset for A/H recognition. Our results show limited performance, suggesting that more adapted multi-modal models are required for accurate A/H recognition. Better methods for modeling spatio-temporal and multimodal fusion are necessary to leverage conflicts within/across modalities.

3D point cloud registration in laparoscopic surgery estimates the transformation between an intraoperative organ reconstructed from video and its preoperative mesh. Because ground-truth transformations are unavailable for real data, supervised networks are trained on synthetic organ pairs. At test time, real reconstructions differ from synthetic data and are noisy, sparse, and occluded, which degrades correspondence estimation. Test-time adaptation (TTA) can reduce this domain shift, but existing methods mainly rely on logits, entropy, class prototypes, or cache memories unavailable in registration. Registration also involves paired inputs with an asymmetric shift that primarily affects the intraoperative cloud. We analyse and modify state-of-the-art TTA methods from three families to 3D registration: model, normalization, and input adaptation. We analyze four representative approaches based on auxiliary-task model updates, backpropagation-free token purging, feature alignment, and layer-normalization calibration. We modify them to handle asymmetric shifts between preoperative and intraoperative point clouds and replace classification-based entropy objectives. Using a correspondence-based model trained on clean synthetic source data, we evaluate adaptation to corrupted synthetic and real target data on P2P and P2ILReg. For synthetic targets, we apply eight corruptions, including uniform noise and global density reduction, at five severity levels. All methods improve registration on P2P, whereas normalization adaptation degrades performance on P2ILReg. Considering the computational overhead of backpropagation-based adaptation, input adaptation is the most promising option for laparoscopic surgery, providing low inference latency and consistent error reductions across datasets. Code: https://github.com/ninaa-git/survey_pc_registration_tta

Personalized facial expression recognition (FER) involves adapting a machine learning model using samples from labeled sources and unlabeled target domains. Given the challenges of recognizing subtle expressions with considerable interpersonal variability, state-of-the-art unsupervised domain adaptation (UDA) methods focus on the multi-source UDA (MSDA) setting, where each domain corresponds to a specific subject, and improve model accuracy and robustness. However, when adapting to a specific target, the diverse nature of multiple source domains translates to a large shift between source and target data. State-of-the-art MSDA methods for FER address this domain shift by considering all the sources to adapt to the target representations. Nevertheless, adapting to a target subject presents significant challenges due to large distributional differences between source and target domains, often resulting in negative transfer. In addition, integrating all sources simultaneously increases computational costs and causes misalignment with the target. To address these issues, we propose a progressive MSDA approach that gradually introduces information from subjects (source domains) based on their similarity to the target subject. This will ensure that only the most relevant sources from the target are selected, which helps avoid the negative transfer caused by dissimilar sources. During adaptation, the source domains are introduced in a curriculum manner. We first exploit the closest sources to reduce the distribution shift with the target and then move towards the furthest while only considering the most relevant sources based on the predetermined threshold. Furthermore, to mitigate catastrophic forgetting caused by the incremental introduction of source subjects, we implemented a density-based memory mechanism that preserves the most relevant historical source samples for adaptation. Our extensive experiments(1) show the effectiveness of our proposed method on challenging FER datasets: Biovid, UNBC-McMaster, Aff-Wild2, and BAH. Further, performance is evaluated on a cross-dataset setting (UNBC-McMaster -> BioVid), showing the importance of gradually adapting to source subjects.

Training fast generators on new domains with limited data remains challenging for two reasons. First, adapting a pretrained diffusion or flow model to a new domain leaves its costly multi-step sampling unaddressed, and existing acceleration methods are tied to the source parameterization–ε, x, v, or u–leaving heterogeneous pretrained models with no common acceleration target. Second, while adversarial refinement is proven effective for few-step quality, it is formulated only for instantaneous-velocity flows, not for the finite-interval average velocities that MeanFlow (MF) models predict. We address both problems. We propose MeanFlow-Transfer, which maps heterogeneous source outputs into a shared velocity representation, uses it to initialize an MF generator from the source weights, and optimizes an MF objective on the target domain. This unifies adaptation and acceleration in a single training loop across a broad range of pretrained models. We then introduce Continuous Adversarial MeanFlow, a post-training stage that extends continuous adversarial flow models from instantaneous velocities to MF's finite-interval average velocities. CAMF contrasts changes in a learned potential between real and predicted interval endpoints, recovering fine detail that MF regression averages away, and reduces to the instantaneous criterion in the vanishing-interval limit. Adapting four ImageNet-based source models–DiT (ε), SiT (v), JiT (x), iMF (u)–to five target domains, MF-T with CAMF matches or exceeds the fine-tuned teacher in FID and FDD at up to 125× fewer Neural Function Evaluations (NFEs), while CAMF improves MF-T's few-step FID by 29% on average.

Recent advances in deep learning (DL) and computational capacity have enabled facial affective behavior analysis (FABA) to progress from static images captured in controlled settings to fine-grained analysis of facial expressions in real world video data. However, training accurate DL models for FABA typically requires large-scale, expert-annotated datasets, which are costly to obtain and inherently noisy due to the ambiguity of labeling subtle facial expressions and action units (AUs). To mitigate these challenges, weakly supervised learning (WSL) has emerged as a promising paradigm for training models with weak annotations. In this paper, we present a structured taxonomy of WSL scenarios for FABA, organized according to the type of weak annotation and the specific affective task. Building on this taxonomy, we provide a critical synthesis of representative WSL methods for both classification (expression and AU recognition) and regression (expression and AU intensity estimation) tasks, focusing on their core methodological ideas, strengths, and limitations. Furthermore, we systematically summarize the comparative performance of WSL approaches along with widely adopted experimental setups and evaluation proto cols. Our critical assessment identifies key challenges and future research directions, including the need for efficient adaptation of foundation models and for the development of robust, scalable FABA systems suitable for real-world applications.

The 10th Affective Behavior Analysis in-the-Wild (ABAW) Workshop and Competition, held at CVPR 2026, continues to advance research on modelling, analysis, understanding of human affect and behavior in real-world, unconstrained environments. The workshop maintains its dual structure, comprising both a competition and a paper track. The ABAW Competition introduces a diverse set of challenges targeting key aspects of affective and behavioral understanding, including continuous affect (valence-arousal) estimation, discrete affect (expression and action unit) recognition, as well as more complex behavior analysis tasks, such as emotional mimicry intensity estimation, ambivalence/hesitancy recognition and fine-grained violence detection. These challenges are built upon large-scale in-the-wild datasets, providing comprehensive benchmarks for state-of-the-art approaches. In parallel, the paper track presents a wide range of contributions spanning pose, motion behavior estimation, affect modelling multimodal learning, benchmarks, datasets evaluation protocols, fairness, robustness deployment. Overall, the 10th ABAW Workshop and Competition continues to serve as a key platform for benchmarking, collaboration and innovation, shaping the development of next-generation multimodal, human-centered AI systems.

In continual learning, the primary challenge is to learn new information without forgetting old knowledge. A common solution addresses this trade-off through regularization, penalizing changes to parameters critical for previous tasks. In most cases, this regularization term is directly added to the training loss and optimized with standard gradient descent, which blends learning and retention signals into a single update and does not explicitly separate essential parameters from redundant ones. As task sequences grow, this coupling can over-constrain the model, limiting forward transfer and leading to inefficient use of capacity. We propose a different approach that separates task learning from stability enforcement via operator splitting. The learning step focuses on minimizing the current task loss, while a proximal stability step applies a sparse regularizer to prune unnecessary parameters and preserve task-relevant ones. This turns the stability-plasticity into a negotiated update between two complementary operators, rather than a conflicting gradient. We provide theoretical justification for the splitting method on the continual-learning objective, and demonstrate that our proposed solver achieves state-of-the-art results on standard benchmarks, improving both stability and adaptability without the need for replay buffers, Bayesian sampling, or meta-learning components.

Transfer learning of diffusion models to smaller target domains is challenging, as naively fine-tuning the model often results in poor generalization. Test-time guidance methods help mitigate this by offering controllable improvements in image fidelity through a trade-off with sample diversity. However, this benefit comes at a high computational cost, typically requiring dual forward passes during sampling. We propose the Domain-guided Fine-tuning (DogFit) method, an effective guidance mechanism for diffusion transfer learning that maintains controllability without incurring additional computational overhead. DogFit injects a domain-aware guidance offset into the training loss, effectively internalizing the guided behavior during the fine-tuning process. The domain-aware design is motivated by our observation that during fine-tuning, the unconditional source model offers a stronger marginal estimate than the target model. To support efficient controllable fidelity–diversity trade-offs at inference, we encode the guidance strength value as an additional model input through a lightweight conditioning mechanism. We further investigate the optimal placement and timing of the guidance offset during training and propose two simple scheduling strategies, i.e., late-start and cut-off, which improve generation quality and training stability. Experiments on DiT and SiT backbones across six diverse target domains show that DogFit can outperform prior guidance methods in transfer learning in terms of FID and FD DINOV2 while requiring up to 2x fewer sampling TFLOPS.

Retrieving user-specified objects from complex scenes remains a challenging task, especially when queries are ambiguous or involve multiple similar objects. Existing open-vocabulary detectors operate in a one-shot manner, lacking the ability to refine predictions based on user feedback. To address this, we propose IntRec, an interactive object retrieval framework that refines predictions based on user feedback. At its core is an Intent State (IS) that maintains dual memory sets for positive anchors (confirmed cues) and negative constraints (rejected hypotheses). A contrastive alignment function ranks candidate objects by maximizing similarity to positive cues while penalizing rejected ones, enabling fine-grained disambiguation in cluttered scenes. Our interactive framework provides substantial improvements in retrieval accuracy without additional supervision. On LVIS, IntRec achieves 35.4 AP, outperforming OVMR, CoDet, and CAKE by +2.3, +3.7, and +0.5, respectively. On the challenging LVIS-Ambiguous benchmark, it improves performance by +7.9 AP over its one-shot baseline after a single corrective feedback, with less than 30 ms of added latency per interaction.