ICCV 2025 论文列表

Self-supervised blind denoising for Poisson-Gaussian noise remains a challenging task. Pseudo-supervised pairs constructed from single noisy images re-corrupt the signal and degrade the performance. The visible blindspots solve the information loss in masked inputs. However, without explicitly noise sensing, mean square error as an objective function cannot adjust denoising intensities for dynamic noise levels, leading to noticeable residual noise. In this paper, we propose Blind2Sound, a simple yet effective approach to overcome residual noise in denoised images. The proposed adaptive re-visible loss senses noise levels and performs personalized denoising without noise residues while retaining the signal lossless. The theoretical analysis of intermediate medium gradients guarantees stable training, while the Cramer Gaussian loss acts as a regularization to facilitate the accurate perception of noise levels and improve the performance of the denoiser. Experiments on synthetic and real-world datasets show the superior performance of our method, especially for single-channel images. The code is publicly available from this link.

Blind video super-resolution (BVSR) is a low-level vision task which aims to generate high-resolution videos from low-resolution counterparts in unknown degradation scenarios. Existing approaches typically predict blur kernels that are spatially invariant in each video frame or even the entire video. These methods do not consider potential spatio-temporal varying degradations in videos, resulting in suboptimal BVSR performance. In this context, we propose a novel BVSR model based on Implicit Kernels, BVSR-IK, which constructs a multi-scale kernel dictionary parameterized by implicit neural representations. It also employs a newly designed recurrent Transformer to predict the coefficient weights for accurate filtering in both frame correction and feature alignment. Experimental results have demonstrated the effectiveness of the proposed BVSR-IK, when compared with four state-of-the-art BVSR models on three commonly used datasets, with BVSR-IK outperforming the second best approach, FMA-Net, by up to 0.59 dB in PSNR. Source code will be available at https://github.com/QZ1-boy/BVSR-IK.

Blind deblurring is an ill-posed inverse problem that involves recovering both the clear image and the blur kernel from a single blurry image. In real photography, longer exposure time results in lots of noise in the blurry image. Although existing blind deblurring methods produce satisfactory results on blurred images with little or no noise, they struggle to handle high noise levels. Strong noise compromises the accuracy of the estimated kernel and significantly reduces the quality of the deblurring results. To address this challenge, we propose a Residual Guidance Strategy (RGS) to suppress the influence of noise. Our method leverages adjacent coarser-scale information in the image pyramid to guide the blur kernel estimation in the current scale. Therefore, for blurred images with unknown noise levels and types, our method still estimates more accurate blur kernels, which are essential for subsequent non-blind restoration. Extensive experiments on both synthetic and real datasets have demonstrated that our method consistently outperforms numerous state-of-the-art methods under high levels of noise quantitatively and qualitatively.

Natural language offers a highly intuitive interface for enabling localized fine-grained edits of 3D shapes. However, prior works face challenges in preserving global coherence while locally modifying the input 3D shape. In this work, we introduce an inpainting-based framework for editing shapes represented as point clouds. Our approach leverages foundation 3D diffusion models for achieving localized shape edits, adding structural guidance in the form of a partial conditional shape, ensuring that other regions correctly preserve the shape's identity. Furthermore, to encourage identity preservation also within the local edited region, we propose an inference-time coordinate blending algorithm which balances reconstruction of the full shape with inpainting at a progression of noise levels during the inference process. Our coordinate blending algorithm seamlessly blends the original shape with its edited version, enabling a fine-grained editing of 3D shapes, all while circumventing the need for computationally expensive and often innacurate inversion. Extensive experiments show that our method outperforms alternative techniques across a wide range of metrics that evaluate both fidelity to the original shape and also adherence to the textual description. We will release our code and trained models.

We propose a novel and general framework to disentangle video data into its dynamic motion and static content components. Our proposed method is a self-supervised pipeline with less assumptions and inductive biases than previous works: it utilizes a transformer-based architecture to jointly generate flexible implicit features for frame-wise motion and clip-wise content, and incorporates a low-bitrate vector quantization as an information bottleneck to promote disentanglement and form a meaningful discrete motion space. The bitrate-controlled latent motion and content are used as conditional inputs to a denoising diffusion model to facilitate self-supervised representation learning. We validate our disentangled representation learning framework on real world talking head videos with motion transfer and auto-regressive motion generation tasks. Furthermore, we also show that our method can generalize to other type of video data, such as pixel sprites of 2D cartoon characters. Our work presents a new perspective on self-supervised learning of disentangled video representations, contributing to the broader field of video analysis and generation.

We present billboard Splatting (BBSplat) - a novel approach for novel view synthesis based on textured geometric primitives. BBSplat represents the scene as a set of optimizable textured planar primitives with learnable RGB textures and alpha-maps to control their shape. BBSplat primitives can be used in any Gaussian Splatting pipeline as drop-in replacements for Gaussians. The proposed primitives close the rendering quality gap between 2D and 3D Gaussian Splatting (GS), enabling the accurate extraction of 3D mesh as in the 2DGS framework. Additionally, the explicit nature of planar primitives enables the use of the ray-tracing effects in rasterization.Our novel regularization term encourages textures to have a sparser structure, enabling an efficient compression that leads to a reduction in the storage space of the model up to x17 times compared to 3DGS. Our experiments show the efficiency of BBSplat on standard datasets of real indoor and outdoor scenes such as Tanks&Temples, DTU, and Mip-NeRF-360. Namely, we achieve a state-of-the-art PSNR of 29.72 for DTU at Full HD resolution.

Open vocabulary Human-Object Interaction (HOI) detection is a challenging task that detects all <human, verb, object> triplets of interest in an image, even those that are not pre-defined in the training set. Existing approaches typically rely on output features generated by large Vision-Language Models (VLMs) to enhance the generalization ability of interaction representations. However, the visual features produced by VLMs are holistic and coarse-grained, which contradicts the nature of detection tasks. To address this issue, we propose a novel Bilateral Collaboration framework for open vocabulary HOI detection (BC-HOI). This framework includes an Attention Bias Guidance (ABG) component, which guides the VLM to produce fine-grained instance-level interaction features according to the attention bias provided by the HOI detector. It also includes a Large Language Model (LLM)-based Supervision Guidance (LSG) component, which provides fine-grained token-level supervision for the HOI detector by the LLM component of the VLM. LSG enhances the ability of ABG to generate high-quality attention bias. We conduct extensive experiments on two popular benchmarks: HICO-DET and V-COCO, consistently achieving superior performance in the open vocabulary and closed settings. The code will be released in Github.

Text-Video Retrieval aims to find the most relevant text (or video) candidate given a video (or text) query from large-scale online databases. Recent work leverages multi-modal large language models (MLLMs) to improve retrieval, especially for long or complex query-candidate pairs. However, we observe that the naive application of MLLMs, i.e., retrieval based on candidate likelihood, introduces candidate prior bias, favoring candidates with inherently higher priors over those more relevant to the query. To this end, we propose a novel retrieval framework, Bidirectional Likelihood Estimation with MLLM (BLiM), which leverages both query and candidate likelihoods by training the model to generate text from a given video as well as video features from a given text. Furthermore, we introduce Candidate Prior Normalization (CPN), a simple yet effective training-free score calibration module designed to mitigate candidate prior bias in candidate likelihood. On four Text-Video Retrieval benchmarks, our BLiM equipped with CPN outperforms previous state-of-the-art models by 6.4 R@1 on average, effectively alleviating candidate prior bias and emphasizing query-candidate relevance. Our in-depth analysis across various multi-modal tasks beyond retrieval highlights the broad applicability of CPN which enhances visual understanding by reducing reliance on textual priors. Code is available at \href https://github.com/mlvlab/BLiM https://github.com/mlvlab/BLiM .

Weakly supervised semantic segmentation (WSSS) aims to generate dense labels using sparse annotations, such as image-level labels. Existing class activation map (CAM) generation methods have been able to locate rough objects. However, due to the limited information provided by image level labels, the bias activation problem, including over-activation, becomes another key obstacle in WSSS. To rectify such bias activation, we attempt to mine pixel level class feature distribution information from the entire dataset. Specifically, we propose to use normalizing flow to model the class feature distribution of all pixels across the entire dataset and design a Bias-Resilient WSSS framework based on Normalizing Flow (BRNF). Normalizing flow has the ability to map complex distributions to normal distributions. Building upon it, we designed an additional Gaussian mixture classifier which classifies pixels from the perspective of feature distributions, providing supplementary information to the conventional MLP based classifier. In addition, we use this distribution to sample low bias features as positive anchors for contrastive learning, thereby encouraging feature optimization toward the correct low-bias direction. Experimental results demonstrate that our method significantly outperforms existing baselines, achieving state-of-the-art performance on WSSS benchmarks. Code will be available at https://github.com/DpDark/BRNF.

Gender bias in vision-language foundation models (VLMs) raises concerns about their safe deployment and is typically evaluated using benchmarks with gender annotations on real-world images. However, as these benchmarks often contain spurious correlations between gender and non-gender features, such as objects and backgrounds, we identify a critical oversight in gender bias evaluation: Do spurious features distort gender bias evaluation? To address this question, we systematically perturb non-gender features across four widely used benchmarks (COCO-gender, FACET, MIAP, and PHASE) and various VLMs to quantify their impact on bias evaluation. Our findings reveal that even minimal perturbations, such as masking just 10% of objects or weakly blurring backgrounds, can dramatically alter bias scores, shifting metrics by up to 175% in generative VLMs and 43% in CLIP variants. This suggests that current bias evaluations often reflect model responses to spurious features rather than gender bias, undermining their reliability. Since creating spurious feature-free benchmarks is fundamentally challenging, we recommend reporting bias metrics alongside feature-sensitivity measurements to enable a more reliable bias assessment.

AI-generated face detectors trained via supervised learning typically rely on synthesized images from specific generators, limiting their generalization to emerging generative techniques. To overcome this limitation, we introduce a self-supervised method based on bi-level optimization. In the inner loop, we pretrain a vision encoder only on photographic face images using a set of linearly weighted pretext tasks: classification of categorical exchangeable image file format (EXIF) tags, ranking of ordinal EXIF tags, and detection of artificial face manipulations. The outer loop then optimizes the relative weights of these pretext tasks to enhance the coarse-grained detection of manipulated faces, serving as a proxy task for identifying AI-generated faces. In doing so, it aligns self-supervised learning more closely with the ultimate goal of AI-generated face detection. Once pretrained, the encoder remains fixed, and AI-generated faces are detected either as anomalies under a Gaussian mixture model fitted to photographic face features or by a lightweight two-layer perceptron serving as a binary classifier. Extensive experiments demonstrate that our detectors significantly outperform existing approaches in both one-class and binary classification settings, exhibiting strong generalization to unseen generators.

The realistic reconstruction of street scenes is critical for developing real-world simulators in autonomous driving. Most existing methods rely on object pose annotations, using these poses to reconstruct dynamic objects and move them during the rendering process. This dependence on high-precision object annotations limits large-scale and extensive scene reconstruction. To address this challenge, we propose Bezier curve Gaussian splatting (BezierGS), which represents the motion trajectories of dynamic objects using learnable Bezier curves. This approach fully leverages the temporal information of dynamic objects and, through learnable curve modeling, automatically corrects pose errors. By introducing additional supervision on dynamic object rendering and inter-curve consistency constraints, we achieve reasonable and accurate separation and reconstruction of scene elements. Extensive experiments on the Waymo Open Dataset and the nuPlan benchmark demonstrate that BezierGS outperforms state-of-the-art alternatives in both dynamic and static scene components reconstruction and novel view synthesis.

Training-free Neural Architecture Search (NAS) has emerged an efficient way to discover high-performing lightweight models with zero-cost proxies (e.g., the activation-based proxies (AZP)). In this paper, we observe a new negative correlation phenomenon that the correlations of the AZP dramatically decrease to be negative with the increasing number of convolutions, significantly degrading the prediction performance of AZP over target architectures. No existing works focus on such negative correlation and its underlying mechanism. To address this, through deep analysis of the architectural characteristics scored by AZP, we propose a series of AZP design principles and reveal the potential reason of the above phenomenon that high non-linearity dramatically degrades the magnitude of AZP score. Those findings show that existing AZP designs do not obey the proposed principles. Finally, grounded in these insights, we propose a simple yet efficient \underline N egative \underline C orrelations-\underline D efied (NCD) method, which utilize stochastic activation masking (SAM) and non-linear rescaling (NIR) to effectively eliminate negative correlation of AZP and significantly improve performance. Extensive experimental results validate the effectiveness and efficiency of our method, outperforming state-of-the-art methods on mainstream 12 search spaces with 4 real-world tasks.

360deg videos have emerged as a promising medium to represent our dynamic visual world. Compared to the "tunnel vision" of standard cameras, their borderless field of view offers a more complete perspective of our surroundings. While existing video models excel at producing standard videos, their ability to generate full panoramic videos remains elusive. In this paper, we investigate the task of video-to-360deg generation: given a perspective video as input, our goal is to generate a full panoramic video that is consistent with the original video. Unlike conventional video generation tasks, the output's field of view is significantly larger, and the model is required to have a deep understanding of both the spatial layout of the scene and the dynamics of objects to maintain spatio-temporal consistency. To address these challenges, we first leverage the abundant 360deg videos available online and develop a high-quality data filtering pipeline to curate pairwise training data. We then carefully design a series of geometry- and motion-aware operations to facilitate the learning process and improve the quality of 360deg video generation. Experimental results demonstrate that our model can generate realistic and coherent 360deg videos from in-the-wild perspective video. In addition, we showcase its potential applications, including video stabilization, camera viewpoint control, and interactive visual question answering.

Embodied Question Answering (EQA) is a challenging task in embodied intelligence that requires agents to dynamically explore 3D environments, actively gather visual information, and perform multi-step reasoning to answer questions. However, current EQA approaches suffer from critical limitations in exploration efficiency, dataset design, and evaluation metrics. Moreover, existing datasets often introduce biases or prior knowledge, leading to disembodied reasoning, while frontier-based exploration strategies struggle in cluttered environments and fail to ensure fine-grained exploration of task-relevant areas. To address these challenges, we construct the EXPloration-awaRe Embodied queStion anSwering Benchmark (EXPRESS-Bench), the largest dataset designed specifically to evaluate both exploration and reasoning capabilities. EXPRESS-Bench consists of 777 exploration trajectories and 2,044 question-trajectory pairs. To improve exploration efficiency, we propose Fine-EQA, a hybrid exploration model that integrates frontier-based and goal-oriented navigation to guide agents toward task-relevant regions more effectively. Additionally, we introduce a novel evaluation metric, Exploration-Answer Consistency (EAC), which ensures faithful assessment by measuring the alignment between answer grounding and exploration reliability. Extensive experimental comparisons with state-of-the-art EQA models demonstrate the effectiveness of our EXPRESS-Bench in advancing embodied exploration and question reasoning.

Masked Image Modeling (MIM) has emerged as a promising approach for Self-Supervised Learning (SSL) of visual representations. However, the out-of-the-box performance of MIMs is typically inferior to competing approaches. Most users cannot afford fine-tuning due to the need for large amounts of data, high GPU consumption, and specialized user knowledge. Therefore, the practical use of MIM representations is limited. In this paper we ask what is the reason for the poor out-of-the-box performance of MIMs. Is it due to weaker features produced by MIM models, or is it due to suboptimal usage? Through detailed analysis, we show that attention in MIMs is spread almost uniformly over many patches, leading to ineffective aggregation by the [cls] token. Based on this insight, we propose Selective aggregation to better capture the rich semantic information retained in patch tokens, which significantly improves the out-of-the-box performance of MIM.

Text-based person search aims to retrieve specific individuals across camera networks using natural language descriptions. However, current benchmarks often exhibit biases towards common actions like walking or standing, neglecting the critical need for identifying abnormal behaviors in real-world scenarios. To meet such demands, we propose a new task, text-based person anomaly search, locating pedestrians engaged in both routine or anomalous activities via text. To enable the training and evaluation of this new task, we construct a large-scale image-text Pedestrian Anomaly Behavior (PAB) benchmark, featuring a broad spectrum of actions, e.g., running, performing, playing soccer, and the corresponding anomalies, e.g., lying, being hit, and falling of the same identity. The training set of PAB comprises 1,013,605 synthesized image-text pairs of both normalities and anomalies, while the test set includes 1,978 real-world image-text pairs. To validate the potential of PAB, we introduce a cross-modal pose-aware framework, which integrates human pose patterns with identity-based hard negative pair sampling. Extensive experiments on the proposed benchmark show that synthetic training data facilitates the fine-grained behavior retrieval, and the proposed pose-aware method arrives at 84.93% recall@1 accuracy, surpassing other competitive methods.

Recent advancements in multimodal large language models (MLLMs) have opened new avenues for video understanding. However, achieving high performance in zero-shot video tasks remains challenging. Traditional video processing methods rely heavily on fine-tuning to capture nuanced spatial-temporal details, which incurs significant data and computation costs. In contrast, training-free approaches, though efficient, often lack robustness in preserving context-rich features across complex video content. To this end, we propose DyTo, a novel dynamic token merging framework for zero-shot video understanding that adaptively optimizes token efficiency while preserving crucial scene details. DyTo integrates hierarchical frame selection and bipartite token merging strategy to dynamically cluster key frames and selectively compress token sequences, striking a balance between computational efficiency with semantic richness. Extensive experiments across multiple benchmarks demonstrate the effectiveness of DyTo. Our method not only sets a new state-of-the-art for zero-shot video understanding when applied to image-trained MLLMs, but also further boosts the performance of models already fine-tuned on video data. Code is available at https://github.com/Jam1ezhang/DYTO.

Large Vision-Language Models (VLMs) generally contain significantly more visual tokens than their textual counterparts, resulting in a considerable computational burden. Recent efforts have been made to tackle this issue by pruning visual tokens early within the language model. Most existing works use attention scores between text and visual tokens to assess the importance of visual tokens. However, in this study, we first analyze the text-visual attention in the language model and find that this score is not an ideal indicator for visual token pruning. Based on the analysis, We propose VisPruner, a plug-and-play method that utilizes visual cues for more effective token pruning in LVLMs. Specifically, we first use visual attention to select a limited number of significant tokens. Then, we remove duplicate tokens from the remaining ones based on their similarity. By retaining diverse tokens alongside the initially selected important tokens, we maximally preserve the visual information of the input image. Experimental results demonstrate that our VisPruner sustains strong performance across various VLM architectures and reduction ratios, significantly outperforming existing methods based on text-visual attention. Notably, without any training, VisPruner can reduce the FLOPs of LLaVA-1.5-7B by 91% and inference latency by 75%, while maintaining comparable performance. Our code is available at https://github.com/Theia-4869/VisPruner.