CVPR 2025 Papers

Video question grounding (VideoQG) requires models to answer the questions and simultaneously infer the relevant video segments to support the answers. However, existing VideoQG methods usually suffer from spurious cross-modal correlations, leading to a failure to identify the dominant visual scenes that align with the intended question. Moreover, vision-language models exhibit unfaithful generalization performance and lack robustness on challenging downstream tasks such as VideoQG. In this work, we propose a novel VideoQG framework named Cross-modal Causal Relation Alignment (CRA), to eliminate spurious correlations and improve the causal consistency between question-answering and video temporal grounding. Our CRA involves three essential components: i) Gaussian Smoothing Grounding (GSG) module for estimating the time interval via cross-modal attention, which is de-noised by an adaptive Gaussian filter, ii) Cross-Modal Alignment (CMA) enhances the performance of weakly supervised VideoQG by leveraging bidirectional contrastive learning between estimated video segments and QA features, iii) Explicit Causal Intervention (ECI) module for multimodal deconfounding, which involves front-door intervention for vision and back-door intervention for language. Extensive experiments on two VideoQG datasets demonstrate the superiority of our CRA in discovering visually grounded content and achieving robust question reasoning. Codes are available at https://github.com/WissingChen/CRA-GQA.

The lack of a large-scale 3D-text corpus has led recent works to distill open-vocabulary knowledge from vision-language models (VLMs). However, these methods typically rely on a single VLM to align the feature spaces of 3D models within a common language space, which limits the potential of 3D models to leverage the diverse spatial and semantic capabilities encapsulated in various foundation models. In this paper, we propose Cross-modal and Uncertainty-aware Agglomeration for Open-vocabulary 3D Scene Understanding dubbed CUA-O3D, the first model to integrate multiple foundation models--such as CLIP, DINOv2, and Stable Diffusion--into 3D scene understanding. We further introduce a deterministic uncertainty estimation to adaptively distill and harmonize the heterogeneous 2D feature embeddings from these models. Our method addresses two key challenges: (1) incorporating semantic priors from VLMs alongside the geometric knowledge of spatially-aware vision foundation models, and (2) using a novel deterministic uncertainty estimation to capture model-specific uncertainties across diverse semantic and geometric sensitivities, helping to reconcile heterogeneous representations during training. Extensive experiments on ScanNetV2 and Matterport3D demonstrate that our method not only advances open-vocabulary segmentation but also achieves robust cross-domain alignment and competitive spatial perception capabilities.

The goal of image cropping is to identify visually appealing crops in an image. Conventional methods are trained on specific datasets and fail to adapt to new requirements. Recent breakthroughs in large vision-language models (VLMs) enable visual in-context learning without explicit training. However, downstream tasks with VLMs remain under explored. In this paper, we propose an effective approach to leverage VLMs for image cropping. First, we propose an efficient prompt retrieval mechanism for image cropping to automate the selection of in-context examples. Second, we introduce an iterative refinement strategy to iteratively enhance the predicted crops. The proposed framework, we refer to as Cropper, is applicable to a wide range of cropping tasks, including free-form cropping, subject-aware cropping, and aspect ratio-aware cropping. Extensive experiments demonstrate that Cropper significantly outperforms state-of-the-art methods across several benchmarks.

Vision-language models (VLMs) have shown remarkable advancements in multimodal reasoning tasks. However, they still often generate inaccurate or irrelevant responses due to issues like hallucinated image understandings or unrefined reasoning paths. To address these challenges, we introduce Critic-V, a novel framework inspired by the Actor-Critic paradigm to boost the reasoning capability of VLMs. This framework decouples the reasoning process and critic process by integrating two independent components: the Reasoner, which generates reasoning paths based on visual and textual inputs, and the Critic, which provides constructive critique to refine these paths. In this approach, the Reasoner generates reasoning responses according to text prompts, which can evolve iteratively as a policy based on feedback from the Critic. This interaction process was theoretically driven by a reinforcement learning framework where the Critic offers natural language critiques instead of scalar rewards, enabling more nuanced feedback to boost the Reasoner's capability on complex reasoning tasks. The Critic model is trained using Direct Preference Optimization (DPO), leveraging a preference dataset of critiques ranked by Rule-based Reward (RBR) to enhance its critic capabilities. Evaluation results show that the Critic-V framework significantly outperforms existing methods, including GPT-4V, on 5 out of 8 benchmarks, especially regarding reasoning accuracy and efficiency. Combining a dynamic text-based policy for the Reasoner and constructive feedback from the preference-optimized Critic enables a more reliable and context-sensitive multimodal reasoning process. Our approach provides a promising solution to enhance the reliability of VLMs, improving their performance in real-world reasoning-heavy multimodal applications such as autonomous driving and embodied intelligence.

Repetitive action counting, which aims to count periodic movements in a video, is valuable for video analysis applications such as fitness monitoring. However, existing methods largely rely on regression networks with limited representational capacity, which hampers their ability to accurately capture variable periodic patterns. Additionally, their supervised learning on narrow, limited training sets leads to overfitting and restricts their ability to generalize across diverse scenarios. To address these challenges, we propose CountLLM, the first large language model (LLM)-based framework that takes video data and periodic text prompts as inputs and outputs the desired counting value. CountLLM leverages the rich clues from explicit textual instructions and the powerful representational capabilities of pre-trained LLMs for repetitive action counting. To effectively guide CountLLM, we develop a periodicity-based structured template for instructions that describes the properties of periodicity and implements a standardized answer format to ensure consistency. Additionally, we propose a progressive multimodal training paradigm to enhance the periodicity-awareness of the LLM. Empirical evaluations on widely recognized benchmarks demonstrate CountLLM's superior performance and generalization, particularly in handling novel and out-of-domain actions that deviate significantly from the training data, offering a promising avenue for repetitive action counting.

Modeling label correlations has always played a pivotal role in multi-label image classification (MLC), attracting significant attention from researchers. However, recent studies have overemphasized co-occurrence relationships among labels, which can lead to overfitting risk on this overemphasis, resulting in suboptimal models. To tackle this problem, we advocate for balancing correlative and discriminative relationships among labels to mitigate the risk of overfitting and enhance model performance. To this end, we propose the Multi-Label Visual Prompt Tuning framework, a novel and parameter-efficient method that groups classes into multiple class subsets according to label co-occurrence and mutual exclusivity relationships, and then models them respectively to balance the two relationships. In this work, since each group contains multiple classes, multiple prompt tokens are adopted within Vision Transformer (ViT) to capture the correlation or discriminative label relationship within each group, and effectively learn correlation or discriminative representations for class subsets. On the other hand, each group contains multiple group-aware visual representations that may correspond to multiple classes, and the mixture of experts (MoE) model can cleverly assign them from the group-aware to the label-aware, adaptively obtaining label-aware representation, which is more conducive to classification. Experiments on multiple benchmark datasets show that our proposed approach achieves competitive results and outperforms SOTA methods on multiple pre-trained models.

Determining the vanishing points (VPs) in a Manhattan world, as a fundamental task in many 3D vision applications, consists of jointly inferring the line-VP association and locating each VP. Existing methods are, however, either sub-optimal solvers or pursuing global optimality at a significant cost of computing time. In contrast to prior works, we introduce convex relaxation techniques to solve this task for the first time. Specifically, we employ a "soft" association scheme, realized via a truncated multi-selection error, that allows for joint estimation of VPs' locations and line-VP associations. This approach leads to a primal problem that can be reformulated into a quadratically constrained quadratic programming (QCQP) problem, which is then relaxed into a convex semidefinite programming (SDP) problem. To solve this SDP problem efficiently, we present a globally optimal outlier-robust iterative solver (called GlobustVP), which independently searches for one VP and its associated lines in each iteration, treating other lines as outliers. After each independent update of all VPs, the mutual orthogonality between the three VPs in a Manhattan world is reinforced via local refinement. Extensive experiments on both synthetic and real-world data demonstrate that GlobustVP achieves a favorable balance between efficiency, robustness, and global optimality compared to previous works. The code is publicly available at github.com/wu-cvgl/GlobustVP.

Multi-center star shape is a prevalent object shape feature, which has proven effective in model-based image segmentation methods. However, the shape field function induced by the multi-center star shape is non-smooth, and directly applying it to the data-driven image segmentation network architecture design may lead to instability in backpropagation. This paper proposes a convex combination star (CCS) shape, possessing multi-center star shape properties, and has the advantage of effectively controlling the shape of the region through a smooth field function. The sufficient condition of the proposed CCS shape can be combined into the image segmentation neural network structure design through the bridge between the variational segmentation model and the activation function of the data-driven method. Taking Segment Anything Model (SAM) and its improved version as backbone networks, we have shown that the segmentation network architecture with CCS shape properties can greatly improve the accuracy of segmentation results.

Manipulation of facial images to meet specific controls such as pose, expression, and lighting, also referred to as face rigging is a complex task in computer vision. Existing methods are limited by their reliance on image datasets, which necessitates individual-specific fine-tuning and limits their ability to retain fine-grained identity and semantic details, reducing practical usability. To overcome these limitations, we introduce ControlFace, a novel face rigging method conditioned on 3DMM renderings that enables flexible, high-fidelity control. ControlFace employs a dual-branch U-Nets: one, referred to as FaceNet, captures identity and fine details, while the other focuses on generation. To enhance control precision, control mixer module encodes the correlated features between the target-aligned control and reference-aligned control, and a novel guidance method, reference control guidance, steers the generation process for better control adherence. By training on a facial video dataset, we fully utilize FaceNet's rich representations while ensuring control adherence. Extensive experiments demonstrate ControlFace's superior performance in identity preservation, and control precision, highlighting its practicality. Code and pre-trained weights will be publicly available.

Multimodal RLHF usually happens after supervised finetuning (SFT) stage to continually improve vision-language models' (VLMs) comprehension. Conventional wisdom holds its superiority over continual SFT during this preference alignment stage. In this paper, we observe that the inherent value of multimodal RLHF lies in its negative supervision, the logit of the rejected responses. We thus propose a novel negative supervised finetuning (nSFT) approach that fully excavates these information resided. Our nSFT disentangles this negative supervision in RLHF paradigm, and continually aligns VLMs with a simple SFT loss. This is more memory efficient than multimodal RLHF where 2 (e.g., DPO) or 4 (e.g., PPO) large VLMs are strictly required. The effectiveness of nSFT is rigorously proved by comparing it with various multimodal RLHF approaches, across different dataset sources, base VLMs and evaluation metrics. Besides, fruitful of ablations are provided to support our hypothesis. Code will be found in https://github.com/Kevinz-code/nSFT/.

The Audio Description (AD) task aims to generate descriptions of visual elements for visually impaired individuals to help them access long-form video contents, like movie. With video feature, text, character bank and context information as inputs, the generated ADs are able to correspond to the characters by name and provide reasonable, contextual descriptions to help audience understand the storyline of movie. To achieve this goal, we propose to leverage pre-trained foundation models through a simple and unified framework to generate ADs with interleaved multimodal sequence as input, termed as Uni-AD. To enhance the alignment of features across various modalities with finer granularity, we introduce a simple and lightweight module that maps video features into the textual feature space. Moreover, we also propose a character-refinement module to provide more precise information by identifying the main characters who play more significant role in the video context. With these unique designs, we further incorporate contextual information and a contrastive loss into our architecture to generate more smooth and contextual ADs. Experiments on the MAD-eval dataset show that Uni-AD can achieve state-of-the-art performance on AD generation, which demonstrates the effectiveness of our approach. Our code is available at: https://github.com/ant-research/UniAD.

Large-scale multimodal representation learning successfully optimizes for zero-shot transfer at test time. Yet the standard pretraining paradigm (contrastive learning on large amounts of image-text data) does not explicitly encourage representations to support few-shot adaptation. In this work, we propose a simple, but carefully designed extension to multimodal pretraining which enables representations to accommodate additional context. Using this objective, we show that vision-language models can be trained to exhibit significantly increased few-shot adaptation: across 21 downstream tasks, we find up to four-fold improvements in test-time sample efficiency, and average few-shot adaptation gains of over 5%, while retaining zero-shot generalization performance across model scales and training durations. In particular, equipped with simple, training-free, metric-based adaptation mechanisms, our representations surpass significantly more complex optimization-based adaptation schemes.

Posterior sampling in high-dimensional spaces using generative models holds significant promise for various applications, including but not limited to inverse problems and guided generation tasks. Generating diverse posterior samples remains expensive, as existing methods require restarting the entire generative process for each new sample. In this work, we propose a posterior sampling approach that simulates Langevin dynamics in the noise space of a pre-trained generative model. By exploiting the mapping between the noise and data spaces which can be provided by distilled flows or consistency models, our method enables seamless exploration of the posterior without the need to re-run the full sampling chain, drastically reducing computational overhead. Theoretically, we prove a guarantee for the proposed noise-space Langevin dynamics to approximate the posterior, assuming that the generative model sufficiently approximates the prior distribution. Our framework is experimentally validated on image restoration tasks involving noisy linear and nonlinear forward operators applied to LSUN-Bedroom (256 x 256) and ImageNet (64 x 64) datasets. The results demonstrate that our approach generates high-fidelity samples with enhanced semantic diversity even under a limited number of function evaluations, offering superior efficiency and performance compared to existing diffusion-based posterior sampling techniques.

In large vision-language models (LVLMs), images serve as inputs that carry a wealth of information. As the idiom "A picture is worth a thousand words" implies, representing a single image in current LVLMs can require hundreds or even thousands of tokens. This results in significant computational costs, which grow quadratically as input image resolution increases, thereby severely impacting the efficiency. Previous approaches have attempted to reduce the number of image tokens either before or within the early layers of LVLMs. However, these strategies inevitably result in the loss of crucial image information. To address this challenge, we conduct an empirical study revealing that all visual tokens are necessary for LVLMs in the shallow layers, and token redundancy progressively increases in the deeper layers.To this end, we propose ViCo, a conical-style visual concentration strategy for LVLMs to boost their efficiency in both training and inference with neglectable performance loss. Specifically, we partition the LVLM into several stages and drop part of the image tokens at the end of each stage with a pre-defined ratio. The dropping is based on a lightweight similarity calculation with a negligible time overhead. Extensive experiments demonstrate that ViCo can achieve over 40% training time reduction and 55% inference FLOPs acceleration on leading LVLMs like LLaVA-NeXT, maintaining comparable multi-modal performance. Besides, ViCo can also serve as a plug-and-play strategy to accelerate inference in a free way, with better performance and lower inference cost than counterparts.

Vision-language models pre-trained at large scale have shown unprecedented adaptability and generalization to downstream tasks. Although its discriminative potential has been widely explored, its reliability and uncertainty are still overlooked. In this work, we investigate the capabilities of CLIP models under the split conformal prediction paradigm, which provides theoretical guarantees to black-box models based on a small, labeled calibration set. In contrast to the main body of literature on conformal predictors in vision classifiers, foundation models exhibit a particular characteristic: they are pre-trained on a one-time basis on an inaccessible source domain, different from the transferred task. This domain drift negatively affects the efficiency of the conformal sets and poses additional challenges. To alleviate this issue, we propose Conf-OT, a transfer learning setting that operates transductive over the combined calibration and query sets. Solving an optimal transport problem, the proposed method bridges the domain gap between pre-training and adaptation without requiring additional data splits but still maintaining coverage guarantees. We comprehensively explore this conformal prediction strategy on a broad span of 15 datasets and three non-conformity scores. Conf-OT provides consistent relative improvements of up to 20% on set efficiency while being 15 times faster than popular transductive approaches.

Composed image retrieval (CIR) is the task of retrieving a target image specified by a query image and a relative text that describes a semantic modification to the query image. Existing methods in CIR struggle to accurately represent the image and the text modification, resulting in subpar performance. To address this limitation, we introduce a CIR framework, \name , trained with a Text Concept-Consistency loss that encourages the representations of noun phrases in the text modification to better attend to the relevant parts of the query image. To support training with this loss function, we also propose a synthetic data generation pipeline that creates training data from existing CIR datasets or unlabeled images. We show that these components together enable stronger performance on CIR tasks, setting a new state-of-the-art in composed image retrieval in both the supervised and zero-shot settings on multiple benchmark datasets, including CIRR and CIRCO. Source code, model checkpoints, and our new datasets are available at https://github.com/mvrl/ConText-CIR.

The feature attribution method reveals the contribution of input variables to the decision-making process to provide an attribution map for explanation. Existing methods grounded on the information bottleneck principle compute information in a specific layer to obtain attributions, compressing the features by injecting noise via a parametric damping ratio. However, the attribution obtained in a specific layer neglects evidence of the decision-making process distributed across layers. In this paper, we introduce a comprehensive information bottleneck (CoIBA), which discovers the relevant information in each targeted layer to explain the decision-making process. Our core idea is applying information bottleneck in multiple targeted layers to estimate the comprehensive information by sharing a parametric damping ratio across the layers. Leveraging this shared ratio complements the over-compressed information to discover the omitted clues of the decision by sharing the relevant information across the targeted layers. We suggest the variational approach to fairly reflect the relevant information of each layer by upper bounding layer-wise information. Therefore, CoIBA guarantees that the discarded activation is unnecessary in every targeted layer to make a decision. The extensive experimental results demonstrate the enhancement in faithfulness of the feature attributions provided by CoIBA.

Attribute detection is crucial for many computer vision tasks, as it enables systems to describe properties such as color, texture, and material. Current approaches often rely on labor-intensive annotation processes which are inherently limited: objects can be described at an arbitrary level of detail (e.g., color vs. color shades), leading to ambiguities when the annotators are not instructed carefully. Furthermore, they operate within a predefined set of attributes, reducing scalability and adaptability to unforeseen downstream applications. We present Compositional Caching (ComCa), a training-free method for open-vocabulary attribute detection that overcomes these constraints. ComCa requires only the list of target attributes and objects as input, using them to populate an auxiliary cache of images by leveraging web-scale databases and Large Language Models to determine attribute-object compatibility. To account for the compositional nature of attributes, cache images receive soft attribute labels. Those are aggregated at inference time based on the similarity between the input and cache images, refining the predictions of underlying Vision-Language Models (VLMs). Importantly, our approach is model-agnostic, compatible with various VLMs. Experiments on public datasets demonstrate that ComCa significantly outperforms zero-shot and cache-based baselines, competing with recent training-based methods, proving that a carefully designed training-free approach can successfully address open-vocabulary attribute detection.

Recent advancements in all-in-one image restoration models have revolutionized the ability to address diverse degradations through a unified framework. However, parameters tied to specific tasks often remain inactive for other tasks, making mixture-of-experts (MoE) architectures a natural extension. Despite this, MoEs often show inconsistent behavior, with some experts unexpectedly generalizing across tasks while others struggle within their intended scope. This hinders leveraging MoEs' computational benefits by bypassing irrelevant experts during inference.We attribute this undesired behavior to the uniform and rigid architecture of traditional MoEs. To address this, we introduce "complexity experts" -- flexible expert blocks with varying computational complexity and receptive fields. A key challenge is assigning tasks to each expert, as degradation complexity is unknown in advance. Thus, we execute tasks with a simple bias toward lower complexity.To our surprise, this preference effectively drives task-specific allocation, assigning tasks to experts with the appropriate complexity. Extensive experiments validate our approach, demonstrating the ability to bypass irrelevant experts during inference while maintaining superior performance. The proposed MoCE-IR model outperforms state-of-the-art methods, affirming its efficiency and practical applicability. The source code and models are publicly available at \href https://eduardzamfir.github.io/moceir/ \texttt eduardzamfir.github.io/MoCE-IR/