ICCV 2025 论文列表

Retrieval-augmented Generation (RAG) enhances Large Language Models (LLMs) by integrating external knowledge to reduce hallucinations and incorporate up-to-date information without retraining. As an essential part of RAG, external knowledge bases are commonly built by extracting structured data from unstructured PDF documents using Optical Character Recognition (OCR). However, given the imperfect prediction of OCR and the inherent non-uniform representation of structured data, knowledge bases inevitably contain various OCR noises. In this paper, we introduce OHRBench, the first benchmark for understanding the cascading impact of OCR on RAG systems. OHRBench includes 8,561 carefully selected unstructured document images from seven real-world RAG application domains, along with 8,498 Q&A pairs derived from multimodal elements in documents, challenging existing OCR solutions used for RAG. To better understand OCR's impact on RAG systems, we identify two primary types of OCR noise: Semantic Noise and Formatting Noise and apply perturbation to generate a set of structured data with varying degrees of each OCR noise. Using OHRBench, we first conduct a comprehensive evaluation of current OCR solutions and reveal that none is competent for constructing high-quality knowledge bases for RAG systems. We then systematically evaluate the impact of these two noise types and demonstrate the trend relationship between the degree of OCR noise and RAG performance. Our OHRBench, including PDF documents, Q&As, and the ground truth structured data are released at: https://github.com/opendatalab/OHR-Bench

Human perception involves decomposing complex multi-object scenes into time-static object appearance (i.e., size, shape, color) and time-varying object motion (i.e., position, velocity, acceleration). For machines to achieve human-like intelligence in real-world interactions, understanding these physical properties of objects is essential, forming the foundation for dynamic video prediction. While recent advancements in object-centric transformers have demonstrated potential in video prediction, they primarily focus on object appearance, often overlooking motion dynamics, which is crucial for modeling dynamic interactions and maintaining temporal consistency in complex environments. To address these limitations, we propose OCK, a dynamic video prediction model leveraging object-centric kinematics and object slots. We introduce a novel component named Object Kinematics that comprises explicit object motions, serving as an additional attribute beyond conventional appearance features to model dynamic scenes. The Object Kinematics are integrated into various OCK mechanisms, enabling spatiotemporal prediction of complex object interactions over long video sequences. Our model demonstrates superior performance in handling complex scenes with intricate object attributes and motions, highlighting its potential for applicability in vision-related dynamics learning tasks.

Understanding the world from multiple perspectives is essential for intelligent systems operating together, where segmenting common objects across different views remains an open problem. We introduce a new approach that re-defines cross-image segmentation by treating it as a mask matching task. Our method consists of: (1) A Mask-Context Encoder that pools dense DINOv2 semantic features to obtain discriminative object-level representations from FastSAM mask candidates, (2) an Ego-Exo Cross-Attention that fuses multi-perspective observations, (3) a Mask Matching contrastive loss that aligns cross-view features in a shared latent space, and (4) a Hard Negative Adjacent Mining strategy to encourage the model to better differentiate between nearby objects. O-MaMa achieves the state of the art in the Ego-Exo4D Correspondences benchmark, obtaining relative gains of +22 % and +76 % in the Ego2Exo and Exo2Ego IoU against the official challenge baselines, and a +13 % and +6 % compared with the SOTA with 1 % of the training parameters.

Suffering from performance bottlenecks in passively detecting high-quality Deepfake images due to the advancement of generative models, proactive perturbations offer a promising approach to disabling Deepfake manipulations by inserting signals into benign images. However, existing proactive perturbation approaches remain unsatisfactory in several aspects: 1) visual degradation due to direct element-wise addition; 2) limited effectiveness against face swapping manipulation; 3) unavoidable reliance on white- and grey-box settings to involve generative models during training. In this study, we analyze the essence of Deepfake face swapping and argue the necessity of protecting source identities rather than target images, and we propose NullSwap, a novel proactive defense approach that cloaks source image identities and nullifies face swapping under a pure black-box scenario. We design an Identity Extraction module to obtain facial identity features from the source image, while a Perturbation Block is then devised to generate identity-guided perturbations accordingly. Meanwhile, a Feature Block extracts shallow-level image features, which are then fused with the perturbation in the Cloaking Block for image reconstruction. Furthermore, to ensure adaptability across different identity extractors in face swapping algorithms, we propose Dynamic Loss Weighting to adaptively balance identity losses. Experiments demonstrate the outstanding ability of our approach to fool various identity recognition models, outperforming state-of-the-art proactive perturbations in preventing face swapping models from generating images with correct source identities.

In this paper, we explore the task of generating expansive outdoor scenes, ranging from castles to high-rises. Unlike indoor scene generation, which has been a primary focus of prior work, outdoor scene generation presents unique challenges, including wide variations in scene heights and the need for a method capable of rapidly producing large landscapes. To address this, we propose an efficient approach that encodes scene chunks as uniform vector sets, offering better compression and performance than the spatially structured latents used in prior methods. Furthermore, we train an explicit outpainting model for unbounded generation, which improves coherence compared to prior resampling-based inpainting schemes while also speeding up generation by eliminating extra diffusion steps. To facilitate this task, we curate NuiScene43, a small but high-quality set of scenes, preprocessed for joint training. Notably, when trained on scenes of varying styles, our model can blend different environments, such as rural houses and city skyscrapers, within the same scene, highlighting the potential of our curation process to leverage heterogeneous scenes for joint training.

Recent advances in multi-modal large language models (MLLMs) have demonstrated strong performance across various domains; however, their ability to comprehend driving scenes remains less proven. The complexity of driving scenarios, which includes multi-view information, poses significant challenges for existing MLLMs. In this paper, we introduce NuPlanQA-Eval, a multi-view, multi-modal evaluation benchmark for driving scene understanding. To further support generalization to multi-view driving scenarios, we also propose NuPlanQA-1M, a large-scale dataset comprising 1M real-world visual question-answering (VQA) pairs. For context-aware analysis of traffic scenes, we categorize our dataset into nine subtasks across three core skills: Road Environment Perception, Spatial Relations Recognition, and Ego-Centric Reasoning. Furthermore, we present BEV-LLM, integrating Bird's-Eye-View (BEV) features from multi-view images into MLLMs. Our evaluation results reveal key challenges that existing MLLMs face in driving scene-specific perception and spatial reasoning from ego-centric perspectives. In contrast, BEV-LLM demonstrates remarkable adaptability to this domain, outperforming other models in six of the nine subtasks. These findings highlight how BEV integration enhances multi-view MLLMs while also identifying key areas that require further refinement for effective adaptation to driving scenes. NuPlanQA is available at https://github.com/sungyeonparkk/NuPlanQA.

In this paper, we analyze the viewpoint stability of foundational models - specifically, their sensitivity to changes in viewpoint- and define instability as significant feature variations resulting from minor changes in viewing angle, leading to generalization gaps in 3D reasoning tasks. We investigate nine foundational models, focusing on their responses to viewpoint changes, including the often-overlooked accidental viewpoints where specific camera orientations obscure an object's true 3D structure. Our methodology enables recognizing and classifying accidental, stable and other viewpoints using feature representations alone, without accessing the actual images. Our findings indicate that while foundation models consistently encode accidental viewpoints, they vary in their interpretation of other viewpoints due to inherent biases, at times leading to object misclassifications based on geometric resemblance. Through quantitative and qualitative evaluations on three downstream tasks - classification, VQA, and 3D reconstruction - we illustrate the impact of viewpoint instability and underscore the importance of feature robustness across diverse viewing conditions.

Is visual information alone sufficient for visual speech recognition (VSR) in challenging real-world scenarios? Humans do not rely solely on visual information for lip-reading but also incorporate additional cues, such as speech-related context and prior knowledge about the task. However, existing methods have largely overlooked such external information in automatic VSR systems. To systematically explore the role of such information for VSR, we introduce the concept of Peripheral Information. We categorize it into three types based on the relevance to the spoken content: (1) Contextual Guidance (e.g., topic or description of speech), (2) Task Expertise (e.g., human prior experience in lip-reading), and (3) Linguistic Perturbation (irrelevant signals processed alongside meaningful information). Considering the disparity that peripheral information provides additional clues with varying significance while visual input serves as the most direct source for VSR, we propose a framework that introduces a hierarchical processing strategy to handle different modalities. With visual-specific adaptation and a dynamic routing mechanism for multi-modal information, our approach reduces the impact of modality conflicts effectively and enables selective utilization of peripheral information with varying relevance. Leveraging readily available peripheral information, our model achieves a WER of 22.03% on LRS3. Further experiments on AVSpeech demonstrate its generalization in real-world scenarios.

Recently, the generation of dynamic 3D objects from a video has shown impressive results. Existing methods directly optimize Gaussians using whole information in frames. However, when dynamic regions are interwoven with static regions within frames, particularly if the static regions account for a large proportion, existing methods often overlook information in dynamic regions and are prone to overfitting on static regions. This leads to producing results with blurry textures. We consider that decoupling dynamic-static features to enhance dynamic representations can alleviate this issue. Thus, we propose a dynamic-static feature decoupling module (DSFD). Along temporal axes, it regards the regions of current frame features that possess significant differences relative to reference frame features as dynamic features. Conversely, the remaining parts are the static features. Then, we acquire decoupled features driven by dynamic features and current frame features. Moreover, to further enhance the dynamic representation of decoupled features from different viewpoints and ensure accurate motion prediction, we design a temporal-spatial similarity fusion module (TSSF). Along spatial axes, it adaptively selects similar information of dynamic regions. Hinging on the above, we construct a novel approach, DS4D. Experimental results verify our method achieves state-of-the-art (SOTA) results in video-to-4D. In addition, the experiments on a real-world scenario dataset demonstrate its effectiveness on the 4D scene. Project page: https://github.com/LiyingCV/DS4D.

Generalizable Image Super-Resolution aims to enhance model generalization capabilities under unknown degradations. To achieve such goal, the models are expected to focus only on image content-related features instead of degradation details (i.e., overfitting degradations).Recently, numerous approaches such as dropout and feature alignment have been proposed to suppress models' natural tendency to overfitting degradations and yields promising results. Nevertheless, these works have assumed that models overfit to all degradation types (e.g., blur, noise), while through careful investigations in this paper, we discover that models predominantly overfit to noise, largely attributable to the distinct degradation pattern in noise compared to other degradation types. In this paper, we propose a targeted feature denoising framework, comprising noise detection and denoising modules. Our approach represents a general solution that can be seamlessly integrated with existing super-resolution models without requiring architectural modifications. Our framework demonstrates superior performance compared to previous regularization-based methods across five traditional benchmark and datasets, encompassing both synthetic and real-world scenarios.

We propose NormalLoc, a novel visual localization method for estimating the 6-DoF pose of a camera using textureless 3D models. Existing methods often rely on color or texture information, limiting their applicability in scenarios where such information is unavailable. NormalLoc addresses this limitation by using rendered normal images generated from surface normals of 3D models to establish a training scheme for both global descriptor computation and matching. This approach enables robust visual localization even when geometric details are limited. Experimental results demonstrate that NormalLoc achieves state-of-the-art performance for visual localization on textureless 3D models, especially in scenarios with limited geometric detail.

Surface normal estimation serves as a cornerstone for a spectrum of computer vision applications. While numerous efforts have been devoted to static image scenarios, ensuring temporal coherence in video-based normal estimation remains a formidable challenge. Instead of merely augmenting existing methods with temporal components, we present NormalCrafter to leverage the inherent temporal priors of Video Diffusion Models (VDMs). We identify the reason for blurry predictions when directly applying VDMs and introduce Semantic Feature Regularization (SFR) to encourage the model to concentrate on geometric details by aligning diffusion features with fine-grained semantic cues. Moreover, we introduce a two-stage training protocol that leverages both latent and pixel space learning to preserve spatial accuracy while maintaining long temporal context. Extensive evaluations demonstrate the efficacy of our method, showcasing a superior performance in generating temporally consistent normal sequences with intricate details from diverse videos. Code and models are publicly available at https://normalcrafter.github.io/.

Anomaly detection in computational pathology aims to identify rare and scarce anomalies where disease-related data are often limited or missing. Existing anomaly detection methods, primarily designed for industrial settings, face limitations in pathology due to computational constraints, diverse tissue structures, and lack of interpretability. To address these challenges, we propose Ano-NAViLa, a Normal and Abnormal pathology knowledge-augmented Vision-Language model for Anomaly detection in pathology images. Ano-NAViLa is built on a pre-trained vision-language model with a lightweight trainable MLP. By incorporating both normal and abnormal pathology knowledge, Ano-NAViLa enhances accuracy and robustness to variability in pathology images and provides interpretability through image-text associations. Evaluated on two lymph node datasets from different organs, Ano-NAViLa achieves the state-of-the-art performance in anomaly detection and localization, outperforming competing models.

High-quality video generation is crucial for many fields, including the film industry and autonomous driving. However, generating videos with spatiotemporal consistencies remains challenging. Current methods typically utilize attention mechanisms or modify noise to achieve consistent videos, neglecting global spatiotemporal information that could help ensure spatial and temporal consistency during video generation. In this paper, we propose the ***NoiseController***, consisting of **Multi-Level Noise Decomposition**, **Multi-Frame Noise Collaboration**, and **Joint Denoising**, to enhance spatiotemporal consistencies in video generation. In multi-level noise decomposition, we first decompose initial noises into scene-level foreground/background noises, capturing distinct motion properties to model multi-view foreground/background variations. Furthermore, each scene-level noise is further decomposed into individual-level shared and residual components. The shared noise preserves consistency, while the residual component maintains diversity. In multi-frame noise collaboration, we introduce an inter-view spatiotemporal collaboration matrix and an intra-view impact collaboration matrix, which captures mutual cross-view effects and historical cross-frame impacts to enhance video quality. The joint denoising contains two parallel denoising U-Nets to remove each scene-level noise, mutually enhancing video generation. We evaluate our ***NoiseController*** on public datasets focusing on video generation and downstream tasks, demonstrating its state-of-the-art performance.

Building on recent advances in Bayesian statistics and image denoising, we propose Noise2Score3D, a fully unsupervised framework for point cloud denoising. Noise2Score3D learns the score function of the underlying point cloud distribution directly from noisy data, eliminating the need for clean data during training. Using Tweedie's formula, our method performs denoising in a single step, avoiding the iterative processes used in existing unsupervised methods, thus improving both accuracy and efficiency. Additionally, we introduce Total Variation for Point Clouds as a denoising quality metric, which allows for the estimation of unknown noise parameters. Experimental results demonstrate that Noise2Score3D achieves state-of-the-art performance on standard benchmarks among unsupervised learning methods in Chamfer distance and point-to-mesh metrics. Noise2Score3D also demonstrates strong generalization ability beyond training datasets. Our method, by addressing the generalization issue and challenge of the absence of clean data in learning-based methods, paves the way for learning-based point cloud denoising methods in real-world applications.

Spike-based imaging, inspired by the human visual system, offers several advantages, including high temporal resolution and low power consumption, but suffers from significant image degradation in low-light conditions due to noise interference. Restoring spike images under such conditions poses a significant challenge, as traditional frame-based or spike-based techniques are ill-suited to handle such severe noise and unique noise characteristics. This paper proposes a novel approach for restoring low-light spike images using noise-modeled diffusion models. By establishing a noise-embedded spike imaging model under low light, we model the forward diffusion process as the degradation of spike images with proportional and residual terms and incorporate deterministic and non-deterministic components with reverse shifting, enabling the model to capture the distinctive spike noise structure. Additionally, we utilize region mask image, dark current map and spike density value as conditions to further guide the restoration process by providing prompts for degradation regions, deterministic parameters and noise intensity, respectively. Experimental results demonstrate that our method significantly outperforms existing spike-based reconstruction and diffusion-based image restoration methods in both quantitative performance and visual quality. The code and dataset are available at https://github.com/BIT-Vision/SpikeDiffusion.

We introduce SPFSplat, an efficient framework for 3D Gaussian splatting from sparse multi-view images, requiring no ground-truth poses during training or inference. It employs a shared feature extraction backbone, enabling simultaneous prediction of 3D Gaussian primitives and camera poses in a canonical space from unposed inputs within a single feed-forward step. Alongside the rendering loss based on estimated novel-view poses, a reprojection loss is integrated to enforce the learning of pixel-aligned Gaussian primitives for enhanced geometric constraints. This pose-free training paradigm and efficient one-step feed-forward design make SPFSplat well-suited for practical applications. Remarkably, despite the absence of pose supervision, SPFSplat achieves state-of-the-art performance in novel view synthesis even under significant viewpoint changes and limited image overlap. It also surpasses recent methods trained with geometry priors in relative pose estimation. Code and trained models are available on our project page: https://ranrhuang.github.io/spfsplat/.

Detection transformers have been applied to human-object interaction (HOI) detection, enhancing the localization and recognition of human-action-object triplets in images. Despite remarkable progress, this study identifies a critical issue--"Toxic Siblings" bias--which hinders the interaction decoder's learning, as numerous similar yet distinct HOI triplets interfere with and even compete against each other both input side and output side to the interaction decoder. This bias arises from high confusion among sibling triplets/categories, where increased similarity paradoxically reduces precision, as one's gain comes at the expense of its toxic sibling's decline. To address this, we propose two novel debiasing learning objectives--"contrastive-then-calibration" and "merge-then-split"--targeting the input and output perspectives, respectively. The former samples sibling-like incorrect HOI triplets and reconstructs them into correct ones, guided by strong positional priors. The latter first learns shared features among sibling categories to distinguish them from other groups, then explicitly refines intra-group differentiation to preserve uniqueness. Experiments show that we significantly outperform both the baseline (+9.18% mAP on HICO-Det) and the state-of-the-art (+3.59% mAP) across various settings. The source code will be made public.

In-context learning (ICL), a type of universal model, demonstrates exceptional generalization across a wide range of tasks without retraining by leveraging task-specific guidance from context, making it particularly effective for the intricate demands of neuroimaging. However, current ICL models, limited to 2D inputs and thus exhibiting suboptimal performance, struggle to extend to 3D inputs due to the high memory demands of ICL. In this regard, we introduce Neuroverse3D, an ICL model capable of performing multiple neuroimaging tasks in 3D (e.g., segmentation, denoising, inpainting). Neuroverse3D overcomes the large memory consumption associated with 3D inputs through adaptive parallel-sequential context processing and a U-shaped fusion strategy, allowing it to handle an unlimited number of context images. Additionally, we propose an optimized loss function to balance multi-task training and enhance focus on anatomical boundaries. Our study incorporates 43,674 3D multi-modal scans from 19 neuroimaging datasets and evaluates Neuroverse3D on 14 diverse tasks using held-out test sets. The results demonstrate that Neuroverse3D significantly outperforms existing ICL models and closely matches task-specific models, enabling flexible adaptation to medical center variations without retraining. The code and model weights will be made publicly available.