CVPR 2025 论文列表

Test-time prompt tuning for vision-language models (VLMs) is getting attention because of their ability to learn with unlabeled data without fine-tuning. Although test-time prompt tuning methods for VLMs can boost accuracy, the resulting models tend to demonstrate poor calibration, which casts doubts on the reliability and trustworthiness of these models. Notably, more attention needs to be devoted to calibrating the test-time prompt tuning in vision-language models. To this end, we propose a new approach, called O-TPT that introduces orthogonality constraints on the textual features corresponding to the learnable prompts for calibrating test-time prompt tuning in VLMs. Towards introducing orthogonality constraints, we make the following contributions. First, we uncover new insights behind the suboptimal calibration performance of existing methods relying on textual feature dispersion. Second, we show that imposing a simple orthogonalization of textual features is a more effective approach towards obtaining textual dispersion. We conduct extensive experiments on various datasets with different backbones and baselines. The results indicate that our method consistently outperforms the prior state of the art in significantly reducing the overall average calibration error. Also, our method surpasses the zero-shot calibration performance on fine-grained classification tasks. Our code is available at https://github.com/ashshaksharifdeen/O-TPT.

Video Large Language Models (Vid-LLMs) have made remarkable advancements in comprehending video content for QA dialogue. However, they struggle to extend this visual understanding to tasks requiring precise temporal localization, known as Video Temporal Grounding (VTG). To address this, we introduce Number-Prompt (NumPro), a novel method that empowers Vid-LLMs to bridge visual comprehension with temporal grounding by adding unique numerical identifiers to each video frame. Treating a video as a sequence of numbered frame images, NumPro transforms VTG into an intuitive process: flipping through manga panels in sequence. This allows Vid-LLMs to "read" event timelines, accurately linking visual content with corresponding temporal information. Our experiments demonstrate that NumPro significantly boosts VTG performance of top-tier Vid-LLMs without additional computational cost. Furthermore, fine-tuning on a NumPro-enhanced dataset defines a new state-of-the-art for VTG, surpassing previous top-performing methods by up to 6.9% in mIoU for moment retrieval and 8.5% in mAP for highlight detection. The code is available at https://github.com/yongliang-wu/NumPro.

Recent studies have shown that large vision-language models (LVLMs) often suffer from the issue of object hallucinations (OH). To mitigate this issue, we introduce an efficient method that edits the model weights based on an unsafe subspace, which we call HalluSpace in this paper. With truthful and hallucinated text prompts accompanying the visual content as inputs, the HalluSpace can be identified by extracting the hallucinated embedding features and removing the truthful representations in LVLMs. By orthogonalizing the model weights, input features will be projected into the Null space of the HalluSpace to reduce OH, based on which we name our method Nullu. We reveal that HalluSpaces generally contain prior information in the large language models (LLMs) applied to build LVLMs, which have been shown as essential causes of OH in previous studies. Therefore, null space projection suppresses the LLMs' priors to filter out the hallucinated features, resulting in contextually accurate outputs. Experiments show that our method can effectively mitigate OH across different LVLM families without extra inference costs and also show strong performance in general LVLM benchmarks. Code is released at https://github.com/Ziwei-Zheng/Nullu.

The knowledge-based visual question answering (KB-VQA) task involves using external knowledge about the image to assist reasoning. Building on the impressive performance of multimodal large language model (MLLM), recent methods have commenced leveraging MLLM as an implicit knowledge base for reasoning. However, the direct employment of MLLM with raw external knowledge might result in reasoning errors due to misdirected knowledge information. Additionally, MLLM may lack fine-grained perception of visual features, which can result in hallucinations during reasoning. To address these challenges, we propose Notes-guided MLLM Reasoning (NoteMR), a novel framework that guides MLLM in better reasoning by utilizing knowledge notes and visual notes. Specifically, we initially obtain explicit knowledge from an external knowledge base. Then, this explicit knowledge, combined with images, is used to assist the MLLM in generating knowledge notes. These notes are designed to filter explicit knowledge and identify relevant internal implicit knowledge within the MLLM. We then identify highly correlated regions between the images and knowledge notes, retaining them as image notes to enhance the model's fine-grained perception, thereby mitigating MLLM induced hallucinations. Finally, both notes are fed into the MLLM, enabling a more comprehensive understanding of the image-question pair and enhancing the model's reasoning capabilities. Our method achieves state-of-the-art performance on the OK-VQA and A-OKVQA datasets, demonstrating its robustness and effectiveness across diverse VQA scenarios.

Vision-Language Models (VLMs) learn a shared feature space for text and images, enabling the comparison of inputs of different modalities. While prior works demonstrated that VLMs organize natural language representations into regular structures encoding composite meanings, it remains unclear if compositional patterns also emerge in the visual embedding space. In this work, we investigate compositionality in the image domain, where the analysis of compositional properties is challenged by noise and sparsity of visual data. We address these problems and propose a framework, called Geodesically Decomposable Embeddings (GDE), that approximates image representations with geometry-aware compositional structures in the latent space. We demonstrate that visual embeddings of pre-trained VLMs exhibit a compositional arrangement, and evaluate the effectiveness of this property in the tasks of compositional classification and group robustness. GDE achieves stronger performance in compositional classification compared to its counterpart method that assumes linear geometry of the latent space. Notably, it is particularly effective for group robustness, where we achieve higher results than task-specific solutions. Our results indicate that VLMs can automatically develop a human-like form of compositional reasoning in the visual domain, making their underlying processes more interpretable. Code is available at https://github.com/BerasiDavide/vlm_image_compositionality.

Current image generation models can effortlessly produce high-quality, highly realistic images, but this also increases the risk of misuse. In various Text-to-Image or Image-to-Image tasks, attackers can generate a series of images containing inappropriate content by simply editing the language modality input. To mitigate this security concern, numerous guarding or defensive strategies have been proposed, with a particular emphasis on safeguarding language modality. However, in practical applications, threats in the vision modality, particularly in tasks involving the editing of real-world images, present heightened security risks as they can easily infringe upon the rights of the image owner. Therefore, this paper employs a method named typographic attack to reveal that various image generation models are also susceptible to threats within the vision modality. Furthermore, we also evaluate the defense performance of various existing methods when facing threats in the vision modality and uncover their ineffectiveness. Finally, we propose the Vision Modal Threats in Image Generation Models (VMT-IGMs) dataset, which would serve as a baseline for evaluating the vision modality vulnerability of various image generation models.

Large language models (LLMs) have significantly enhanced cross-modal understanding capabilities by integrating visual encoders with textual embeddings, giving rise to multimodal large language models (MLLMs). However, these models struggle with non-natural images such as geometric and charts, particularly in fields like education and finance. Despite efforts to collect datasets and fine-tune the MLLMs, the gap with natural image understanding is still evident, and the cost of collecting large and diverse non-natural image datasets is high. To address this, we analyzed the limitations of transformer-based vision encoders(ViT) within existing MLLMs from a frequency perspective. Studies have shown that ViT models are less effective at capturing high-frequency information, impairing their ability to capture elements like points, lines, and angles in non-natural images. In response, we introduced FM-ViT, a frequency-modulated vision encoder that utilizes Fourier decomposition to extract high and low frequency components from self-attention features and re-weight them during tuning to non-natural images. In addition, we combine the features of CNN models with FM-ViT and propose EDGE, an MLLM with enhanced graphical encoders tailored for understanding non-natural images. Extensive experiments have confirmed the effectiveness of our FM-ViT and EDGE in 4 types.

Noise synthesis is a promising solution for addressing the data shortage problem in data-driven low-light RAW image denoising. However, accurate noise synthesis methods often necessitate labor-intensive calibration and profiling procedures during preparation, preventing them from landing to practice at scale. This work introduces a practically simple noise synthesis pipeline based on detailed analyses of noise properties and extensive justification of widespread techniques. Compared to other approaches, our proposed pipeline eliminates the cumbersome system gain calibration and signal-independent noise profiling steps, reducing the preparation time for noise synthesis from days to hours. Meanwhile, our method exhibits strong denoising performance, showing an up to 0.54dB PSNR improvement over the current state-of-the-art noise synthesis technique.

Diffusion models have achieved impressive success in generating photorealistic images, but challenges remain in ensuring precise semantic alignment with input prompts. Optimizing the initial noisy latent offers a more efficient alternative to modifying model architectures or prompt engineering for improving semantic alignment. A latest approach, InitNo, refines the initial noisy latent by leveraging attention maps; however, these maps capture only limited information, and the effectiveness of InitNo is highly dependent on the initial starting point, as it tends to converge on a local optimum near this point. To this end, this paper proposes leveraging the language comprehension capabilities of large vision-language models (LVLMs) to guide the optimization of the initial noisy latent, and introduces the Noise Diffusion process, which updates the noisy latent to generate semantically faithful images while preserving distribution consistency. Furthermore, we provide a theoretical analysis of the condition under which the update improves semantic faithfulness. Experimental results demonstrate the effectiveness and adaptability of our framework, consistently enhancing semantic alignment across various diffusion models.

Scanning Transmission Electron Microscopy (STEM) enables the observation of atomic arrangements at sub-angstrom resolution, allowing for atomically resolved analysis of the physical and chemical properties of materials. However, due to the effects of noise, electron beam damage, sample thickness, etc, obtaining satisfactory atomic-level images is often challenging. Enhancing STEM images can reveal clearer structural details of materials. Nonetheless, existing STEM image enhancement methods usually overlook unique features in the frequency domain, and existing datasets lack realism and generality. To resolve these issues, in this paper, we develop noise calibration, data synthesis, and enhancement methods for STEM images. We first present a STEM noise calibration method, which is used to synthesize more realistic STEM images. The parameters of background noise, scan noise, and pointwise noise are obtained by statistical analysis and fitting of real STEM images containing atoms. Then we use these parameters to develop a more general dataset that considers both regular and random atomic arrangements and includes both HAADF and BF mode images. Finally, we design a spatial-frequency interactive network for STEM image enhancement, which can explore the information in the frequency domain formed by the periodicity of atomic arrangement. Experimental results show that our data is closer to real STEM images and achieves better enhancement performances together with our network. Code will be available at https://github.com/HeasonLee/SFIN.

Over the last decade, many notable methods have emerged to tackle the computational resource challenge of the high resolution image recognition (HRIR). They typically focus on identifying and aggregating a few salient regions for classification, discarding sub-salient areas for low training consumption. Nevertheless, many HRIR tasks necessitate the exploration of wider regions to model objects and contexts, which limits their performance in such scenarios. To address this issue, we present a DBPS strategy to enable training with more patches at low consumption. Specifically, in addition to a fundamental buffer that stores the embeddings of most salient patches, DBPS further employs an auxiliary buffer to recycle those sub-salient ones. To reduce the computational cost associated with gradients of sub-salient patches, these patches are primarily used in the forward pass to provide sufficient information for classification. Meanwhile, only the gradients of the salient patches are back-propagated to update the entire network. Moreover, we design a Multiple Instance Learning (MIL) architecture that leverages aggregated information from salient patches to filter out uninformative background within sub-salient patches for better accuracy. Besides, we introduce the random patch drop to accelerate training process and uncover informative regions. Experiment results demonstrate the superiority of our method in terms of both accuracy and training consumption against other advanced methods. The code is available in the https://github.com/Qinrong-NKU/DBPS.

Obtaining high-quality, practically usable 3D models of biological plants remains a significant challenge in computer vision and graphics. In this paper, we present a novel method for generating realistic 3D plant models from single-view photographs. Our approach employs a neural decomposition technique to learn a lightweight hierarchical box representation from the image, effectively capturing the structures and botanical features of plants. Then, this representation can be subsequently refined through a shape-guided parametric modeling module to produce complete 3D plant models. By combining hierarchical learning and parametric modeling, our method generates structured 3D plant assets with fine geometric details. Notably, through learning the decomposition in different levels of detail, our method can adapt to two distinct plant categories: outdoor trees and houseplants, each with unique appearance features. Within the scope of plant modeling, our method is the first comprehensive solution capable of reconstructing both plant categories from single-view images.

In scene graph generation (SGG), the accurate prediction of unseen triples is essential for its effectiveness in downstream vision-language tasks. We hypothesize that the predicates of unseen triples can be viewed as transformations of seen predicates in feature space, and the essence of the zero-shot task is to bridge the gap caused by this transformation. Traditional models, however, have difficulty addressing this challenge, which we attribute to their inability to model the predicates equivariant. To overcome this limitation, we introduce a novel framework based on capsule networks (CAPSGG). We propose a Three-Stream Pipeline that generates modality-specific representations for predicates, while building low-level predicate capsules of these modalities. Then these capsules are aggregated into high-level predicate capsules using a Routing Capsule Layer. In addition, we introduce GroupLoss to aggregate capsules with the same predicate label into groups. This replaces the global loss with the intra-group loss, effectively balancing the learning of predicate invariance and equivariant features, while mitigating the impact of the severe long-tail distribution of the predicate categories. Our extensive experiments demonstrate the notable superiority of our approach over state-of-the-art methods, with zero-shot indicators outperforming up to 132.26% on SGCls task than the T-CAR [21]. Our code will be available upon publication.

In recent text-video retrieval, the use of additional captions from vision-language models has shown promising effects on the performance. However, existing models using additional captions often have struggled to capture the rich semantics, including temporal changes, inherent in the video. In addition, incorrect information caused by generative models can lead to inaccurate retrieval. To address these issues, we propose a new framework, Narrating the Video (NarVid), which strategically leverages the comprehensive information available from frame-level captions, the narration. The proposed NarVid exploits narration in multiple ways: 1) feature enhancement through cross-modal interactions between narration and video, 2) query-aware adaptive filtering to suppress irrelevant or incorrect information, 3) dual-modal matching score by adding query-video similarity and query-narration similarity, and 4) hard-negative loss to learn discriminative features from multiple perspectives using the two similarities from different views. Experimental results demonstrate that NarVid achieves state-of-the-art performance on various benchmark datasets.

We release NSD-Imagery, a benchmark dataset of human fMRI activity paired with mental images, to complement the existing Natural Scenes Dataset (NSD), a large-scale dataset of fMRI activity paired with seen images that enabled unprecedented improvements in fMRI-to-image reconstruction efforts. Recent models trained on NSD have been evaluated only on seen image reconstruction. Using NSD-Imagery, it is possible to assess how well these models perform on mental image reconstruction. This is a challenging generalization requirement because mental images are encoded in human brain activity with relatively lower signal-to-noise and spatial resolution; however, generalization from seen to mental imagery is critical for real-world applications in medical domains and brain-computer interfaces, where the desired information is always internally generated. We provide benchmarks for a suite of recent NSD-trained open-source visual decoding models (MindEye1, MindEye2, Brain Diffuser, iCNN, Takagi et al.) on NSD-Imagery, and show that the performance of decoding methods on mental images is largely decoupled from performance on vision reconstruction. We further demonstrate that architectural choices significantly impact cross-decoding performance: models employing simple linear decoding architectures and multimodal feature decoding generalize better to mental imagery, while complex architectures tend to overfit visual training data. Our findings indicate that mental imagery datasets are critical for the development of practical applications, and establish NSD-Imagery as a useful resource for better aligning visual decoding methods with this goal.

The emergence of vision-language foundation models, such as CLIP, has revolutionized image-text representation, enabling a broad range of applications via prompt learning. Despite its promise, real-world datasets often contain noisy labels that can degrade prompt learning performance. In this paper, we demonstrate that using mean absolute error (MAE) loss in prompt learning, named PromptMAE, significantly enhances robustness against noisy labels while maintaining high accuracy. Though MAE is straightforward and recognized for its robustness, it is rarely used in noisy-label learning due to its slow convergence and poor performance outside prompt learning scenarios. To elucidate the robustness of PromptMAE, we leverage feature learning theory to show that MAE can suppress the influence of noisy samples, thereby improving the signal-to-noise ratio and enhancing overall robustness. Additionally, we introduce PromptOT, a prompt-based optimal transport data purification method to enhance the robustness further. PromptOT employs text encoder representations in vision-language models as prototypes to construct an optimal transportation matrix. This matrix effectively partitions datasets into clean and noisy subsets, allowing for the application of cross-entropy loss to the clean subset and MAE loss to the noisy subset. Our Noise-Label Prompt Learning method, named NLPrompt, offers a simple and efficient approach that leverages the expressive representation and precise alignment capabilities of vision-language models for robust prompt learning. We validate NLPrompt through extensive experiments across various noise settings, demonstrating significant performance improvements.

We introduce a novel method for pre-training of large-scale vision encoders. Building on recent advancements in autoregressive pre-training of vision models, we extend this framework to a multimodal setting, i.e., images and text. In this paper, we present AIMV2, a family of generalist vision encoders characterized by a straightforward pre-training process, scalability, and remarkable performance across a range of downstream tasks. This is achieved by pairing the vision encoder with a multimodal decoder that autoregressively generates raw image patches and text tokens. Our encoders excel not only in multimodal evaluations but also in vision benchmarks such as localization, grounding, and classification. Notably, our AIMV2-3B encoder achieves 89.5% accuracy on ImageNet-1k with a frozen trunk. Fur- thermore, AIMV2 consistently outperforms state-of-the-art contrastive models (e.g., CLIP, SigLIP) in multimodal im- age understanding across diverse settings.

Efficiently retrieving and synthesizing information from large-scale multimodal collections has become a critical challenge. However, existing video retrieval datasets suffer from scope limitations, primarily focusing on matching descriptive but vague queries with small collections of professionally edited, English-centric videos. To address this gap, we introduce MultiVENT 2.0, a large-scale, multilingual event-centric video retrieval benchmark featuring a collection of more than 218,000 news videos and over 3,900 queries targeting specific world events. These queries specifically target information found in the visual content, audio, embedded text, and text metadata of the videos, requiring systems leverage all these sources to succeed at the task. Preliminary results show that state-of-the-art vision-language models struggle significantly with this task, and while alternative approaches show promise, they are still insufficient to adequately address this problem. These findings underscore the need for more robust multimodal retrieval systems, as effective video retrieval is a crucial step towards multimodal content understanding and generation.