Yunhui Guo

Weiguo Pian, Shentong Mo, Yunhui Guo et al.

In this paper, we introduce audio-visual class-incremental learning, a class-incremental learning scenario for audio-visual video recognition. We demonstrate that joint audio-visual modeling can improve class-incremental learning, but current methods fail to preserve semantic similarity between audio and visual features as incremental step grows. Furthermore, we observe that audio-visual correlations learned in previous tasks can be forgotten as incremental steps progress, leading to poor performance. To overcome these challenges, we propose AV-CIL, which incorporates Dual-Audio-Visual Similarity Constraint (D-AVSC) to maintain both instance-aware and class-aware semantic similarity between audio-visual modalities and Visual Attention Distillation (VAD) to retain previously learned audio-guided visual attentive ability. We create three audio-visual class-incremental datasets, AVE-Class-Incremental (AVE-CI), Kinetics-Sounds-Class-Incremental (K-S-CI), and VGGSound100-Class-Incremental (VS100-CI) based on the AVE, Kinetics-Sounds, and VGGSound datasets, respectively. Our experiments on AVE-CI, K-S-CI, and VS100-CI demonstrate that AV-CIL significantly outperforms existing class-incremental learning methods in audio-visual class-incremental learning. Code and data are available at: https://github.com/weiguoPian/AV-CIL_ICCV2023.

Tsung-Wei Ke, Jyh-Jing Hwang, Yunhui Guo et al.

Unsupervised semantic segmentation aims to discover groupings within and across images that capture object and view-invariance of a category without external supervision. Grouping naturally has levels of granularity, creating ambiguity in unsupervised segmentation. Existing methods avoid this ambiguity and treat it as a factor outside modeling, whereas we embrace it and desire hierarchical grouping consistency for unsupervised segmentation. We approach unsupervised segmentation as a pixel-wise feature learning problem. Our idea is that a good representation shall reveal not just a particular level of grouping, but any level of grouping in a consistent and predictable manner. We enforce spatial consistency of grouping and bootstrap feature learning with co-segmentation among multiple views of the same image, and enforce semantic consistency across the grouping hierarchy with clustering transformers between coarse- and fine-grained features. We deliver the first data-driven unsupervised hierarchical semantic segmentation method called Hierarchical Segment Grouping (HSG). Capturing visual similarity and statistical co-occurrences, HSG also outperforms existing unsupervised segmentation methods by a large margin on five major object- and scene-centric benchmarks. Our code is publicly available at https://github.com/twke18/HSG .

Xiaofan Yu, Yunhui Guo, Sicun Gao et al.

Unsupervised lifelong learning refers to the ability to learn over time while memorizing previous patterns without supervision. Although great progress has been made in this direction, existing work often assumes strong prior knowledge about the incoming data (e.g., knowing the class boundaries), which can be impossible to obtain in complex and unpredictable environments. In this paper, motivated by real-world scenarios, we propose a more practical problem setting called online self-supervised lifelong learning without prior knowledge. The proposed setting is challenging due to the non-iid and single-pass data, the absence of external supervision, and no prior knowledge. To address the challenges, we propose Self-Supervised ContrAstive Lifelong LEarning without Prior Knowledge (SCALE) which can extract and memorize representations on the fly purely from the data continuum. SCALE is designed around three major components: a pseudo-supervised contrastive loss, a self-supervised forgetting loss, and an online memory update for uniform subset selection. All three components are designed to work collaboratively to maximize learning performance. We perform comprehensive experiments of SCALE under iid and four non-iid data streams. The results show that SCALE outperforms the state-of-the-art algorithm in all settings with improvements up to 3.83%, 2.77% and 5.86% in terms of kNN accuracy on CIFAR-10, CIFAR-100, and TinyImageNet datasets.

Yue Zhang, Liqiang Jing, Jia Li et al.

Multimodal Large Language Models have achieved strong performance in single-video understanding, yet their ability to reason across multiple videos remains limited. Existing approaches typically concatenate multiple videos into a single input and perform direct inference, which introduces training-inference mismatch, information loss from frame compression, and a lack of explicit cross-video coordination. Meanwhile, current multi-video benchmarks primarily emphasize event-level comparison, leaving identity-level matching, fine-grained discrimination, and structured multi-step reasoning underexplored. To address these gaps, we introduce MVX-Bench, a Multi-Video Cross-Dimension Benchmark that reformulates 11 classical computer vision tasks into a unified multi-video question-answering framework, comprising 1,442 questions over 4,255 videos from diverse real-world datasets. We further propose SAMA, a Skill-Augmented Agentic Framework for Multi-Video Understanding, which integrates visual tools, task-specific skills, and a conflict-aware verification mechanism to enable iterative and structured reasoning. Experimental results show that SAMA outperforms strong open-source baselines and GPT on MVX-Bench, and ablations validate the effectiveness of skill design and conflict resolution.

Zhihang Ren, Jefferson Ortega, Yifan Wang et al. · berkeley

Human affect recognition has been a significant topic in psychophysics and computer vision. However, the currently published datasets have many limitations. For example, most datasets contain frames that contain only information about facial expressions. Due to the limitations of previous datasets, it is very hard to either understand the mechanisms for affect recognition of humans or generalize well on common cases for computer vision models trained on those datasets. In this work, we introduce a brand new large dataset, the Video-based Emotion and Affect Tracking in Context Dataset (VEATIC), that can conquer the limitations of the previous datasets. VEATIC has 124 video clips from Hollywood movies, documentaries, and home videos with continuous valence and arousal ratings of each frame via real-time annotation. Along with the dataset, we propose a new computer vision task to infer the affect of the selected character via both context and character information in each video frame. Additionally, we propose a simple model to benchmark this new computer vision task. We also compare the performance of the pretrained model using our dataset with other similar datasets. Experiments show the competing results of our pretrained model via VEATIC, indicating the generalizability of VEATIC. Our dataset is available at https://veatic.github.io.

Haoliang Wang, Chen Zhao, Yunhui Guo et al.

Two prevalent types of distributional shifts in machine learning are the covariate shift (as observed across different domains) and the semantic shift (as seen across different classes). Traditional OOD detection techniques typically address only one of these shifts. However, real-world testing environments often present a combination of both covariate and semantic shifts. In this study, we introduce a novel problem, semantic OOD detection across domains, which simultaneously addresses both distributional shifts. To this end, we introduce two regularization strategies: domain generalization regularization, which ensures semantic invariance across domains to counteract the covariate shift, and OOD detection regularization, designed to enhance OOD detection capabilities against the semantic shift through energy bounding. Through rigorous testing on three standard domain generalization benchmarks, our proposed framework showcases its superiority over conventional domain generalization approaches in terms of OOD detection performance. Moreover, it holds its ground by maintaining comparable InD classification accuracy.

Yangxiao Lu, Ninad Khargonkar, Zesheng Xu et al.

We introduce a novel robotic system for improving unseen object instance segmentation in the real world by leveraging long-term robot interaction with objects. Previous approaches either grasp or push an object and then obtain the segmentation mask of the grasped or pushed object after one action. Instead, our system defers the decision on segmenting objects after a sequence of robot pushing actions. By applying multi-object tracking and video object segmentation on the images collected via robot pushing, our system can generate segmentation masks of all the objects in these images in a self-supervised way. These include images where objects are very close to each other, and segmentation errors usually occur on these images for existing object segmentation networks. We demonstrate the usefulness of our system by fine-tuning segmentation networks trained on synthetic data with real-world data collected by our system. We show that, after fine-tuning, the segmentation accuracy of the networks is significantly improved both in the same domain and across different domains. In addition, we verify that the fine-tuned networks improve top-down robotic grasping of unseen objects in the real world.

Yunhui Guo, Youren Zhang, Yubei Chen et al.

Given an image set without any labels, our goal is to train a model that maps each image to a point in a feature space such that, not only proximity indicates visual similarity, but where it is located directly encodes how prototypical the image is according to the dataset. Our key insight is to perform unsupervised feature learning in hyperbolic instead of Euclidean space, where the distance between points still reflect image similarity, and yet we gain additional capacity for representing prototypicality with the location of the point: The closer it is to the origin, the more prototypical it is. The latter property is simply emergent from optimizing the usual metric learning objective: The image similar to many training instances is best placed at the center of corresponding points in Euclidean space, but closer to the origin in hyperbolic space. We propose an unsupervised feature learning algorithm in Hyperbolic space with sphere pACKing. HACK first generates uniformly packed particles in the Poincaré ball of hyperbolic space and then assigns each image uniquely to each particle. Images after congealing are regarded more typical of the dataset it belongs to. With our feature mapper simply trained to spread out training instances in hyperbolic space, we observe that images move closer to the origin with congealing, validating our idea of unsupervised prototypicality discovery. We demonstrate that our data-driven prototypicality provides an easy and superior unsupervised instance selection to reduce sample complexity, increase model generalization with atypical instances and robustness with typical ones.

Wenjie Zhao, Jia Li, Xin Dong et al.

Semantic segmentation models, while effective for in-distribution categories, face challenges in real-world deployment due to encountering out-of-distribution (OoD) objects. Detecting these OoD objects is crucial for safety-critical applications. Existing methods rely on anomaly scores, but choosing a suitable threshold for generating masks presents difficulties and can lead to fragmentation and inaccuracy. This paper introduces a method to convert anomaly \textbf{S}core \textbf{T}o segmentation \textbf{M}ask, called S2M, a simple and effective framework for OoD detection in semantic segmentation. Unlike assigning anomaly scores to pixels, S2M directly segments the entire OoD object. By transforming anomaly scores into prompts for a promptable segmentation model, S2M eliminates the need for threshold selection. Extensive experiments demonstrate that S2M outperforms the state-of-the-art by approximately 20% in IoU and 40% in mean F1 score, on average, across various benchmarks including Fishyscapes, Segment-Me-If-You-Can, and RoadAnomaly datasets.

Weiguo Pian, Yiyang Nan, Shijian Deng et al.

In this paper, we introduce a novel continual audio-visual sound separation task, aiming to continuously separate sound sources for new classes while preserving performance on previously learned classes, with the aid of visual guidance. This problem is crucial for practical visually guided auditory perception as it can significantly enhance the adaptability and robustness of audio-visual sound separation models, making them more applicable for real-world scenarios where encountering new sound sources is commonplace. The task is inherently challenging as our models must not only effectively utilize information from both modalities in current tasks but also preserve their cross-modal association in old tasks to mitigate catastrophic forgetting during audio-visual continual learning. To address these challenges, we propose a novel approach named ContAV-Sep (\textbf{Cont}inual \textbf{A}udio-\textbf{V}isual Sound \textbf{Sep}aration). ContAV-Sep presents a novel Cross-modal Similarity Distillation Constraint (CrossSDC) to uphold the cross-modal semantic similarity through incremental tasks and retain previously acquired knowledge of semantic similarity in old models, mitigating the risk of catastrophic forgetting. The CrossSDC can seamlessly integrate into the training process of different audio-visual sound separation frameworks. Experiments demonstrate that ContAV-Sep can effectively mitigate catastrophic forgetting and achieve significantly better performance compared to other continual learning baselines for audio-visual sound separation. Code is available at: \url{https://github.com/weiguoPian/ContAV-Sep_NeurIPS2024}.

Yuhang Liu, Wenjie Zhao, Yunhui Guo

Task Incremental Learning (TIL) is a specialized form of Continual Learning (CL) in which a model incrementally learns from non-stationary data streams. Existing TIL methodologies operate under the closed-world assumption, presuming that incoming data remains in-distribution (ID). However, in an open-world setting, incoming samples may originate from out-of-distribution (OOD) sources, with their task identities inherently unknown. Continually detecting OOD samples presents several challenges for current OOD detection methods: reliance on model outputs leads to excessive dependence on model performance, selecting suitable thresholds is difficult, hindering real-world deployment, and binary ID/OOD classification fails to provide task-level identification. To address these issues, we propose a novel continual OOD detection method called the Hierarchical Two-sample Tests (H2ST). H2ST eliminates the need for threshold selection through hypothesis testing and utilizes feature maps to better exploit model capabilities without excessive dependence on model performance. The proposed hierarchical architecture enables task-level detection with superior performance and lower overhead compared to non-hierarchical classifier two-sample tests. Extensive experiments and analysis validate the effectiveness of H2ST in open-world TIL scenarios and its superiority to the existing methods. Code is available at \href{https://github.com/YuhangLiuu/H2ST}{https://github.com/YuhangLiuu/H2ST}.

Sarthak Kumar Maharana, Baoming Zhang, Leonid Karlinsky et al.

Although open-vocabulary classification models like Contrastive Language Image Pretraining (CLIP) have demonstrated strong zero-shot learning capabilities, their robustness to common image corruptions remains poorly understood. Through extensive experiments, we show that zero-shot CLIP lacks robustness to common image corruptions during test-time, necessitating the adaptation of CLIP to unlabeled corrupted images using test-time adaptation (TTA). However, we found that existing TTA methods have severe limitations in adapting CLIP due to their unimodal nature. To address these limitations, we propose $\texttt{BATCLIP}$, a bimodal $\textbf{online}$ TTA method designed to improve CLIP's robustness to common image corruptions. The key insight of our approach is not only to adapt the visual encoders for improving image features but also to strengthen the alignment between image and text features by promoting a stronger association between the image class prototype, computed using pseudo-labels, and the corresponding text feature. We evaluate our approach on benchmark image corruption datasets and achieve state-of-the-art results in online TTA for CLIP. Furthermore, we evaluate our proposed TTA approach on various domain generalization datasets to demonstrate its generalization capabilities. Our code is available at https://github.com/sarthaxxxxx/BATCLIP

Yuxuan Li, Sarthak Kumar Maharana, Yunhui Guo

With the increasing prevalence of Machine Learning as a Service (MLaaS) platforms, there is a growing focus on deep neural network (DNN) watermarking techniques. These methods are used to facilitate the verification of ownership for a target DNN model to protect intellectual property. One of the most widely employed watermarking techniques involves embedding a trigger set into the source model. Unfortunately, existing methodologies based on trigger sets are still susceptible to functionality-stealing attacks, potentially enabling adversaries to steal the functionality of the source model without a reliable means of verifying ownership. In this paper, we first introduce a novel perspective on trigger set-based watermarking methods from a feature learning perspective. Specifically, we demonstrate that by selecting data exhibiting multiple features, also referred to as \emph{multi-view data}, it becomes feasible to effectively defend functionality stealing attacks. Based on this perspective, we introduce a novel watermarking technique based on Multi-view dATa, called MAT, for efficiently embedding watermarks within DNNs. This approach involves constructing a trigger set with multi-view data and incorporating a simple feature-based regularization method for training the source model. We validate our method across various benchmarks and demonstrate its efficacy in defending against model extraction attacks, surpassing relevant baselines by a significant margin. The code is available at: \href{https://github.com/liyuxuan-github/MAT}{https://github.com/liyuxuan-github/MAT}.

Ouyang Xu, Baoming Zhang, Ruiyu Mao et al.

Deep learning models for visual recognition often exhibit systematic errors due to underrepresented semantic subpopulations. Although existing debugging frameworks can pinpoint these failures by identifying key failure attributes, repairing the model effectively remains difficult. Current solutions often rely on manually designed prompts to generate synthetic training images -- an approach prone to distribution shift and semantic errors. To overcome these challenges, we introduce a model repair module that builds on an interpretable failure attribution pipeline. Our approach uses a conditional text-to-image model to generate semantically faithful and targeted images for failure cases. To preserve the quality and relevance of the generated samples, we further employ a large vision-language model (LVLM) to filter the outputs, enforcing alignment with the original data distribution and maintaining semantic consistency. By retraining vision models with this rare-case-augmented synthetic dataset, we significantly reduce errors associated with rare cases. Our experiments demonstrate that this targeted repair strategy improves model robustness without introducing new bugs. Code is available at https://github.com/oxu2/SafeFix

Sarthak Kumar Maharana, Saksham Singh Kushwaha, Baoming Zhang et al.

While recent audio-visual models have demonstrated impressive performance, their robustness to distributional shifts at test-time remains not fully understood. Existing robustness benchmarks mainly focus on single modalities, making them insufficient for thoroughly assessing the robustness of audio-visual models. Motivated by real-world scenarios where shifts can occur $\textit{simultaneously}$ in both audio and visual modalities, we introduce $\texttt{AVROBUSTBENCH}$, a comprehensive benchmark designed to evaluate the test-time robustness of audio-visual recognition models. $\texttt{AVROBUSTBENCH}$ comprises four audio-visual benchmark datasets, $\texttt{AUDIOSET-2C}$, $\texttt{VGGSOUND-2C}$, $\texttt{KINETICS-2C}$, and $\texttt{EPICKITCHENS-2C}$, each incorporating 75 bimodal audio-visual corruptions that are $\textit{co-occurring}$ and $\textit{correlated}$. Through extensive evaluations, we observe that state-of-the-art supervised and self-supervised audio-visual models exhibit declining robustness as corruption severity increases. Furthermore, online test-time adaptation (TTA) methods, on $\texttt{VGGSOUND-2C}$ and $\texttt{KINETICS-2C}$, offer minimal improvements in performance under bimodal corruptions. We further propose $\texttt{AV2C}$, a simple TTA approach enabling on-the-fly cross-modal fusion by penalizing high-entropy samples, which achieves improvements on $\texttt{VGGSOUND-2C}$. We hope that $\texttt{AVROBUSTBENCH}$ will steer the development of more effective and robust audio-visual TTA approaches. Our code is available $\href{https://github.com/sarthaxxxxx/AV-C-Robustness-Benchmark}{here}$.

Wenjie Zhao, Jia Li, Xin Dong et al.

Open-set test-time adaptation (OSTTA) addresses the challenge of adapting models to new environments where out-of-distribution (OOD) samples coexist with in-distribution (ID) samples affected by distribution shifts. In such settings, covariate shift-for example, changes in weather conditions such as snow-can alter ID samples, reducing model reliability. Consequently, models must not only correctly classify covariate-shifted ID (csID) samples but also effectively reject covariate-shifted OOD (csOOD) samples. Entropy minimization is a common strategy in test-time adaptation to maintain ID performance under distribution shifts, while entropy maximization is widely applied to enhance OOD detection. Several studies have sought to combine these objectives to tackle the challenges of OSTTA. However, the intrinsic conflict between entropy minimization and maximization inevitably leads to a trade-off between csID classification and csOOD detection. In this paper, we first analyze the limitations of entropy maximization in OSTTA and then introduce an angular loss to regulate feature norm magnitudes, along with a feature-norm loss to suppress csOOD logits, thereby improving OOD detection. These objectives form ROSETTA, a $\underline{r}$obust $\underline{o}$pen-$\underline{se}$t $\underline{t}$est-$\underline{t}$ime $\underline{a}$daptation. Our method achieves strong OOD detection while maintaining high ID classification performance on CIFAR-10-C, CIFAR-100-C, Tiny-ImageNet-C and ImageNet-C. Furthermore, experiments on the Cityscapes validate the method's effectiveness in real-world semantic segmentation, and results on the HAC dataset demonstrate its applicability across different open-set TTA setups.

Yuhang Liu, Wenjie Zhao, Yunhui Guo

Selecting an appropriate pre-trained source model is a critical, yet computationally expensive, task in transfer learning. Model Transferability Estimation (MTE) methods address this by providing efficient proxy metrics to rank models without full fine-tuning. In practice, the choice of which MTE metric to use is often ad hoc or guided simply by a metric's average historical performance. However, we observe that the effectiveness of MTE metrics is highly task-dependent and no single metric is universally optimal across all target datasets. To address this gap, we introduce MetaRank, a meta-learning framework for automatic, task-aware MTE metric selection. We formulate metric selection as a learning-to-rank problem. Rather than relying on conventional meta-features, MetaRank encodes textual descriptions of both datasets and MTE metrics using a pretrained language model, embedding them into a shared semantic space. A meta-predictor is then trained offline on diverse meta-tasks to learn the intricate relationship between dataset characteristics and metric mechanisms, optimized with a listwise objective that prioritizes correctly ranking the top-performing metrics. During the subsequent online phase, MetaRank efficiently ranks the candidate MTE metrics for a new, unseen target dataset based on its textual description, enabling practitioners to select the most appropriate metric a priori. Extensive experiments across 11 pretrained models and 11 target datasets demonstrate the strong effectiveness of our approach.

Wenjie Zhao, Jia Li, Mingrui Liu et al.

``How long can I live and remain free of cancer?'' is often the first question a patient asks after receiving a cancer diagnosis and treatment. Accurate survival prediction helps alleviate psychological distress and supports risk stratification and personalized treatment planning. Recent survival prediction frameworks have shown strong performance using computed tomography (CT) images. However, variations in imaging acquisition introduce out-of-distribution (OOD) samples caused by covariate shifts that undermine model reliability. Despite this challenge, to our knowledge, no existing benchmark systematically studies OOD detection in cancer survival prediction. To address this gap, we introduce the Cancer sURvival bEnchmark for OOD Detection (CURE-OOD), the first benchmark for systematically evaluating OOD detection in survival prediction under controlled acquisition-induced distribution shifts. CURE-OOD defines scanner-parameter-based training, in-distribution (ID), and OOD test splits across four survival prediction tasks. Our experiments show that covariate shifts notably reduce survival prediction performance. It also shows that mainstream classification-oriented OOD detectors can fail in survival prediction. Finally, we include HazardDev as a simple survival-aware reference baseline for OOD detection. CURE-OOD enables systematic analysis of how distribution shifts affect both downstream survival performance and OOD detectability.

Qifan Zhang, Yunhui Guo, Yu Xiang

We introduce the problem of continual distillation learning (CDL) in order to use knowledge distillation (KD) to improve prompt-based continual learning (CL) models. The CDL problem is valuable to study since the use of a larger vision transformer (ViT) leads to better performance in prompt-based continual learning. The distillation of knowledge from a large ViT to a small ViT improves the inference efficiency for prompt-based CL models. We empirically found that existing KD methods such as logit distillation and feature distillation cannot effectively improve the student model in the CDL setup. To address this issue, we introduce a novel method named Knowledge Distillation based on Prompts (KDP), in which globally accessible prompts specifically designed for knowledge distillation are inserted into the frozen ViT backbone of the student model. We demonstrate that our KDP method effectively enhances the distillation performance in comparison to existing KD methods in the CDL setup.

Ruiyu Mao, Ouyang Xu, Yunhui Guo

Active learning is a commonly used approach that reduces the labeling effort required to train deep neural networks. However, the effectiveness of current active learning methods is limited by their closed-world assumptions, which assume that all data in the unlabeled pool comes from a set of predefined known classes. This assumption is often not valid in practical situations, as there may be unknown classes in the unlabeled data, leading to the active open-set annotation problem. The presence of unknown classes in the data can significantly impact the performance of existing active learning methods due to the uncertainty they introduce. To address this issue, we propose a novel data-centric active learning method called NEAT that actively annotates open-set data. NEAT is designed to label known classes data from a pool of both known and unknown classes unlabeled data. It utilizes the clusterability of labels to identify the known classes from the unlabeled pool and selects informative samples from those classes based on a consistency criterion that measures inconsistencies between model predictions and local feature distribution. Unlike the recently proposed learning-centric method for the same problem, NEAT is much more computationally efficient and is a data-centric active open-set annotation method. Our experiments demonstrate that NEAT achieves significantly better performance than state-of-the-art active learning methods for active open-set annotation.

Zilong Wang, Zhiyang Dou, Yuan Liu et al.

In this paper, we present WonderHuman to reconstruct dynamic human avatars from a monocular video for high-fidelity novel view synthesis. Previous dynamic human avatar reconstruction methods typically require the input video to have full coverage of the observed human body. However, in daily practice, one typically has access to limited viewpoints, such as monocular front-view videos, making it a cumbersome task for previous methods to reconstruct the unseen parts of the human avatar. To tackle the issue, we present WonderHuman, which leverages 2D generative diffusion model priors to achieve high-quality, photorealistic reconstructions of dynamic human avatars from monocular videos, including accurate rendering of unseen body parts. Our approach introduces a Dual-Space Optimization technique, applying Score Distillation Sampling (SDS) in both canonical and observation spaces to ensure visual consistency and enhance realism in dynamic human reconstruction. Additionally, we present a View Selection strategy and Pose Feature Injection to enforce the consistency between SDS predictions and observed data, ensuring pose-dependent effects and higher fidelity in the reconstructed avatar. In the experiments, our method achieves SOTA performance in producing photorealistic renderings from the given monocular video, particularly for those challenging unseen parts. The project page and source code can be found at https://wyiguanw.github.io/WonderHuman/.

Jia Li, Wenjie Zhao, Ziru Huang et al.

Unlike traditional visual segmentation, audio-visual segmentation (AVS) requires the model not only to identify and segment objects but also to determine whether they are sound sources. Recent AVS approaches, leveraging transformer architectures and powerful foundation models like SAM, have achieved impressive performance on standard benchmarks. Yet, an important question remains: Do these models genuinely integrate audio-visual cues to segment sounding objects? In this paper, we systematically investigate this issue in the context of robust AVS. Our study reveals a fundamental bias in current methods: they tend to generate segmentation masks based predominantly on visual salience, irrespective of the audio context. This bias results in unreliable predictions when sounds are absent or irrelevant. To address this challenge, we introduce AVSBench-Robust, a comprehensive benchmark incorporating diverse negative audio scenarios including silence, ambient noise, and off-screen sounds. We also propose a simple yet effective approach combining balanced training with negative samples and classifier-guided similarity learning. Our extensive experiments show that state-of-theart AVS methods consistently fail under negative audio conditions, demonstrating the prevalence of visual bias. In contrast, our approach achieves remarkable improvements in both standard metrics and robustness measures, maintaining near-perfect false positive rates while preserving highquality segmentation performance.

Sarthak Kumar Maharana, Baoming Zhang, Yunhui Guo

Real-world vision models in dynamic environments face rapid shifts in domain distributions, leading to decreased recognition performance. Using unlabeled test data, continuous test-time adaptation (CTTA) directly adjusts a pre-trained source discriminative model to these changing domains. A highly effective CTTA method involves applying layer-wise adaptive learning rates for selectively adapting pre-trained layers. However, it suffers from the poor estimation of domain shift and the inaccuracies arising from the pseudo-labels. This work aims to overcome these limitations by identifying layers for adaptation via quantifying model prediction uncertainty without relying on pseudo-labels. We utilize the magnitude of gradients as a metric, calculated by backpropagating the KL divergence between the softmax output and a uniform distribution, to select layers for further adaptation. Subsequently, for the parameters exclusively belonging to these selected layers, with the remaining ones frozen, we evaluate their sensitivity to approximate the domain shift and adjust their learning rates accordingly. We conduct extensive image classification experiments on CIFAR-10C, CIFAR-100C, and ImageNet-C, demonstrating the superior efficacy of our method compared to prior approaches.

Weiguo Pian, Saksham Singh Kushwaha, Zhimin Chen et al.

In this paper, we propose Universal Holistic Audio Generation (UniHAGen), a task for synthesizing comprehensive auditory scenes that include both on-screen and off-screen sounds across diverse domains (e.g., ambient events, musical instruments, and human speech). Prior video-conditioned audio generation models typically focus on producing on-screen environmental sounds that correspond to visible sounding events, neglecting off-screen auditory events. While recent holistic joint text-video-to-audio generation models aim to produce auditory scenes with both on- and off-screen sound but they are limited to non-speech sounds, lacking the ability to generate or integrate human speech. To overcome these limitations, we introduce OmniSonic, a flow-matching-based diffusion framework jointly conditioned on video and text. It features a TriAttn-DiT architecture that performs three cross-attention operations to process on-screen environmental sound, off-screen environmental sound, and speech conditions simultaneously, with a Mixture-of-Experts (MoE) gating mechanism that adaptively balances their contributions during generation. Furthermore, we construct UniHAGen-Bench, a new benchmark with over one thousand samples covering three representative on/off-screen speech-environment scenarios. Extensive experiments show that OmniSonic consistently outperforms state-of-the-art approaches on both objective metrics and human evaluations, establishing a strong baseline for universal and holistic audio generation. Project page: https://weiguopian.github.io/OmniSonic_webpage/

Yue Zhang, Jilei Sun, Yunhui Guo et al.

Video Large Multimodal Models (VLMMs) have made impressive strides in understanding video content, but they often struggle with abstract and adaptive reasoning-the ability to revise their interpretations when new information emerges. In reality, conclusions are rarely set in stone; additional context can strengthen or weaken an initial inference. To address this, we introduce Defeasible Video Entailment (DVidE), a new task that challenges models to think like doubters, constantly updating their reasoning based on evolving evidence. In DVidE, given a video premise and a textual hypothesis, models must determine whether a new update strengthens or weakens the hypothesis (classification version) or generate a coherent update that modifies the entailment relationship (generation version). For solving the classification task, we propose the Chain of Counterfactual Thought framework, utilizing counterfactual reasoning, ASR-enhanced video content, and rationale refinement to reduce inference bias. For the generation task, we develop a framework that combines ASR output with a Large Language Model (LLM) to produce coherent, contextually relevant updates aligned with the intended strengthener or weakener goals. Additionally, we introduce a novel benchmark dataset, with strengthener/weakener annotations and an LLM-based evaluation metric specifically designed for assessing generative performance. Experimental results demonstrate significant improvements, highlighting our proposed method in enhancing dynamic reasoning capabilities of VLMMs.

Yangxiao Lu, Ruosen Li, Liqiang Jing et al.

Visual grounding focuses on detecting objects from images based on language expressions. Recent Large Vision-Language Models (LVLMs) have significantly advanced visual grounding performance by training large models with large-scale datasets. However, the problem remains challenging, especially when similar objects appear in the input image. For example, an LVLM may not be able to differentiate Diet Coke and regular Coke in an image. In this case, if additional reference images of Diet Coke and regular Coke are available, it can help the visual grounding of similar objects. In this work, we introduce a new task named Multimodal Reference Visual Grounding (MRVG). In this task, a model has access to a set of reference images of objects in a database. Based on these reference images and a language expression, the model is required to detect a target object from a query image. We first introduce a new dataset to study the MRVG problem. Then we introduce a novel method, named MRVG-Net, to solve this visual grounding problem. We show that by efficiently using reference images with few-shot object detection and using Large Language Models (LLMs) for object matching, our method achieves superior visual grounding performance compared to the state-of-the-art LVLMs such as Qwen2.5-VL-72B. Our approach bridges the gap between few-shot detection and visual grounding, unlocking new capabilities for visual understanding, which has wide applications in robotics. Project page with our video, code, and dataset: https://irvlutd.github.io/MultiGrounding

Weiguo Pian, Shijian Deng, Shentong Mo et al.

In this paper, we introduce Modality-Inconsistent Continual Learning (MICL), a new continual learning scenario for Multimodal Large Language Models (MLLMs) that involves tasks with inconsistent modalities (image, audio, or video) and varying task types (captioning or question-answering). Unlike existing vision-only or modality-incremental settings, MICL combines modality and task type shifts, both of which drive catastrophic forgetting. To address these challenges, we propose MoInCL, which employs a Pseudo Targets Generation Module to mitigate forgetting caused by task type shifts in previously seen modalities. It also incorporates Instruction-based Knowledge Distillation to preserve the model's ability to handle previously learned modalities when new ones are introduced. We benchmark MICL using a total of six tasks and conduct experiments to validate the effectiveness of our proposed MoInCL. The experimental results highlight the superiority of MoInCL, showing significant improvements over representative and state-of-the-art continual learning baselines.

Wenjie Zhao, Jia Li, Yunhui Guo

Out-of-distribution (OoD) detection and segmentation have attracted growing attention as concerns about AI security rise. Conventional OoD detection methods identify the existence of OoD objects but lack spatial localization, limiting their usefulness in downstream tasks. OoD segmentation addresses this limitation by localizing anomalous objects at pixel-level granularity. This capability is crucial for safety-critical applications such as autonomous driving, where perception modules must not only detect but also precisely segment OoD objects, enabling targeted control actions and enhancing overall system robustness. In this survey, we group current OoD segmentation approaches into four categories: (i) test-time OoD segmentation, (ii) outlier exposure for supervised training, (iii) reconstruction-based methods, (iv) and approaches that leverage powerful models. We systematically review recent advances in OoD segmentation for autonomous-driving scenarios, identify emerging challenges, and discuss promising future research directions.

Yuxuan Li, Yunhui Guo

Architecture plays an important role in deciding the performance of deep neural networks. However, the search for the optimal architecture is often hindered by the vast search space, making it a time-intensive process. Recently, a novel approach known as training-free neural architecture search (NAS) has emerged, aiming to discover the ideal architecture without necessitating extensive training. Training-free NAS leverages various indicators for architecture selection, including metrics such as the count of linear regions, the density of per-sample losses, and the stability of the finite-width Neural Tangent Kernel (NTK) matrix. Despite the competitive empirical performance of current training-free NAS techniques, they suffer from certain limitations, including inconsistent performance and a lack of deep understanding. In this paper, we introduce GradAlign, a simple yet effective method designed for inferring model performance without the need for training. At its core, GradAlign quantifies the extent of conflicts within per-sample gradients during initialization, as substantial conflicts hinder model convergence and ultimately result in worse performance. We evaluate GradAlign against established training-free NAS methods using standard NAS benchmarks, showing a better overall performance. Moreover, we show that the widely adopted metric of linear region count may not suffice as a dependable criterion for selecting network architectures during at initialization.

Yunhui Guo

Transferability estimation has emerged as an important problem in transfer learning. A transferability estimation method takes as inputs a set of pre-trained models and decides which pre-trained model can deliver the best transfer learning performance. Existing methods tackle this problem by analyzing the output of the pre-trained model or by comparing the pre-trained model with a probe model trained on the target dataset. However, neither is sufficient to provide reliable and efficient transferability estimations. In this paper, we present a novel perspective and introduce Kite, as a Kernel-based Improved Transferability Estimation method. Kite is based on the key observations that the separability of the pre-trained features and the similarity of the pre-trained features to random features are two important factors for estimating transferability. Inspired by kernel methods, Kite adopts centered kernel alignment as an effective way to assess feature separability and feature similarity. Kite is easy to interpret, fast to compute, and robust to the target dataset size. We evaluate the performance of Kite on a recently introduced large-scale model selection benchmark. The benchmark contains 8 source dataset, 6 target datasets and 4 architectures with a total of 32 pre-trained models. Extensive results show that Kite outperforms existing methods by a large margin for transferability estimation.

Yunhui Guo, Haoran Guo, Stella Yu

Hyperbolic space can naturally embed hierarchies that often exist in real-world data and semantics. While high-dimensional hyperbolic embeddings lead to better representations, most hyperbolic models utilize low-dimensional embeddings, due to non-trivial optimization and visualization of high-dimensional hyperbolic data. We propose CO-SNE, which extends the Euclidean space visualization tool, t-SNE, to hyperbolic space. Like t-SNE, it converts distances between data points to joint probabilities and tries to minimize the Kullback-Leibler divergence between the joint probabilities of high-dimensional data $X$ and low-dimensional embedding $Y$. However, unlike Euclidean space, hyperbolic space is inhomogeneous: A volume could contain a lot more points at a location far from the origin. CO-SNE thus uses hyperbolic normal distributions for $X$ and hyperbolic \underline{C}auchy instead of t-SNE's Student's t-distribution for $Y$, and it additionally seeks to preserve $X$'s individual distances to the \underline{O}rigin in $Y$. We apply CO-SNE to naturally hyperbolic data and supervisedly learned hyperbolic features. Our results demonstrate that CO-SNE deflates high-dimensional hyperbolic data into a low-dimensional space without losing their hyperbolic characteristics, significantly outperforming popular visualization tools such as PCA, t-SNE, UMAP, and HoroPCA which is also designed for hyperbolic data.

Yunhui Guo, Xudong Wang, Yubei Chen et al.

Hyperbolic space can naturally embed hierarchies, unlike Euclidean space. Hyperbolic Neural Networks (HNNs) exploit such representational power by lifting Euclidean features into hyperbolic space for classification, outperforming Euclidean neural networks (ENNs) on datasets with known semantic hierarchies. However, HNNs underperform ENNs on standard benchmarks without clear hierarchies, greatly restricting HNNs' applicability in practice. Our key insight is that HNNs' poorer general classification performance results from vanishing gradients during backpropagation, caused by their hybrid architecture connecting Euclidean features to a hyperbolic classifier. We propose an effective solution by simply clipping the Euclidean feature magnitude while training HNNs. Our experiments demonstrate that clipped HNNs become super-hyperbolic classifiers: They are not only consistently better than HNNs which already outperform ENNs on hierarchical data, but also on-par with ENNs on MNIST, CIFAR10, CIFAR100 and ImageNet benchmarks, with better adversarial robustness and out-of-distribution detection.

Yunhui Guo, Noel C. Codella, Leonid Karlinsky et al.

Recent progress on few-shot learning largely relies on annotated data for meta-learning: base classes sampled from the same domain as the novel classes. However, in many applications, collecting data for meta-learning is infeasible or impossible. This leads to the cross-domain few-shot learning problem, where there is a large shift between base and novel class domains. While investigations of the cross-domain few-shot scenario exist, these works are limited to natural images that still contain a high degree of visual similarity. No work yet exists that examines few-shot learning across different imaging methods seen in real world scenarios, such as aerial and medical imaging. In this paper, we propose the Broader Study of Cross-Domain Few-Shot Learning (BSCD-FSL) benchmark, consisting of image data from a diverse assortment of image acquisition methods. This includes natural images, such as crop disease images, but additionally those that present with an increasing dissimilarity to natural images, such as satellite images, dermatology images, and radiology images. Extensive experiments on the proposed benchmark are performed to evaluate state-of-art meta-learning approaches, transfer learning approaches, and newer methods for cross-domain few-shot learning. The results demonstrate that state-of-art meta-learning methods are surprisingly outperformed by earlier meta-learning approaches, and all meta-learning methods underperform in relation to simple fine-tuning by 12.8% average accuracy. Performance gains previously observed with methods specialized for cross-domain few-shot learning vanish in this more challenging benchmark. Finally, accuracy of all methods tend to correlate with dataset similarity to natural images, verifying the value of the benchmark to better represent the diversity of data seen in practice and guiding future research.

Yunhui Guo, Yandong Li, Liqiang Wang et al.

There is an increasing number of pre-trained deep neural network models. However, it is still unclear how to effectively use these models for a new task. Transfer learning, which aims to transfer knowledge from source tasks to a target task, is an effective solution to this problem. Fine-tuning is a popular transfer learning technique for deep neural networks where a few rounds of training are applied to the parameters of a pre-trained model to adapt them to a new task. Despite its popularity, in this paper, we show that fine-tuning suffers from several drawbacks. We propose an adaptive fine-tuning approach, called AdaFilter, which selects only a part of the convolutional filters in the pre-trained model to optimize on a per-example basis. We use a recurrent gated network to selectively fine-tune convolutional filters based on the activations of the previous layer. We experiment with 7 public image classification datasets and the results show that AdaFilter can reduce the average classification error of the standard fine-tuning by 2.54%.

Yunhui Guo, Mingrui Liu, Tianbao Yang et al.

Current deep neural networks can achieve remarkable performance on a single task. However, when the deep neural network is continually trained on a sequence of tasks, it seems to gradually forget the previous learned knowledge. This phenomenon is referred to as \textit{catastrophic forgetting} and motivates the field called lifelong learning. Recently, episodic memory based approaches such as GEM \cite{lopez2017gradient} and A-GEM \cite{chaudhry2018efficient} have shown remarkable performance. In this paper, we provide the first unified view of episodic memory based approaches from an optimization's perspective. This view leads to two improved schemes for episodic memory based lifelong learning, called MEGA-I and MEGA-II. MEGA-I and MEGA-II modulate the balance between old tasks and the new task by integrating the current gradient with the gradient computed on the episodic memory. Notably, we show that GEM and A-GEM are degenerate cases of MEGA-I and MEGA-II which consistently put the same emphasis on the current task, regardless of how the loss changes over time. Our proposed schemes address this issue by using novel loss-balancing updating rules, which drastically improve the performance over GEM and A-GEM. Extensive experimental results show that the proposed schemes significantly advance the state-of-the-art on four commonly used lifelong learning benchmarks, reducing the error by up to 18\%.

Yunhui Guo, Yandong Li, Rogerio Feris et al.

There is a growing interest in designing models that can deal with images from different visual domains. If there exists a universal structure in different visual domains that can be captured via a common parameterization, then we can use a single model for all domains rather than one model per domain. A model aware of the relationships between different domains can also be trained to work on new domains with less resources. However, to identify the reusable structure in a model is not easy. In this paper, we propose a multi-domain learning architecture based on depthwise separable convolution. The proposed approach is based on the assumption that images from different domains share cross-channel correlations but have domain-specific spatial correlations. The proposed model is compact and has minimal overhead when being applied to new domains. Additionally, we introduce a gating mechanism to promote soft sharing between different domains. We evaluate our approach on Visual Decathlon Challenge, a benchmark for testing the ability of multi-domain models. The experiments show that our approach can achieve the highest score while only requiring 50% of the parameters compared with the state-of-the-art approaches.

Yunhui Guo, Honghui Shi, Abhishek Kumar et al.

Transfer learning, which allows a source task to affect the inductive bias of the target task, is widely used in computer vision. The typical way of conducting transfer learning with deep neural networks is to fine-tune a model pre-trained on the source task using data from the target task. In this paper, we propose an adaptive fine-tuning approach, called SpotTune, which finds the optimal fine-tuning strategy per instance for the target data. In SpotTune, given an image from the target task, a policy network is used to make routing decisions on whether to pass the image through the fine-tuned layers or the pre-trained layers. We conduct extensive experiments to demonstrate the effectiveness of the proposed approach. Our method outperforms the traditional fine-tuning approach on 12 out of 14 standard datasets.We also compare SpotTune with other state-of-the-art fine-tuning strategies, showing superior performance. On the Visual Decathlon datasets, our method achieves the highest score across the board without bells and whistles.

Yunhui Guo

Deep neural networks are the state-of-the-art methods for many real-world tasks, such as computer vision, natural language processing and speech recognition. For all its popularity, deep neural networks are also criticized for consuming a lot of memory and draining battery life of devices during training and inference. This makes it hard to deploy these models on mobile or embedded devices which have tight resource constraints. Quantization is recognized as one of the most effective approaches to satisfy the extreme memory requirements that deep neural network models demand. Instead of adopting 32-bit floating point format to represent weights, quantized representations store weights using more compact formats such as integers or even binary numbers. Despite a possible degradation in predictive performance, quantization provides a potential solution to greatly reduce the model size and the energy consumption. In this survey, we give a thorough review of different aspects of quantized neural networks. Current challenges and trends of quantized neural networks are also discussed.